From 49f2fc443887fba0aeddf216bc966d0d47fbc483 Mon Sep 17 00:00:00 2001 From: zhangshaozhi Date: Sun, 17 May 2026 20:26:16 +0800 Subject: [PATCH 01/36] fix: make application runnable in production - Dockerfile: include README.md in COPY statement - main.py: add static file serving for built frontend (SPA routing support) - App.vue: wrap template with Naive UI message/dialog/notification providers Co-Authored-By: Claude Sonnet 4.6 --- Dockerfile | 2 +- protoforge/main.py | 10 ++++++++++ web/src/App.vue | 8 +++++++- 3 files changed, 18 insertions(+), 2 deletions(-) diff --git a/Dockerfile b/Dockerfile index babbf5f..2528ade 100644 --- a/Dockerfile +++ b/Dockerfile @@ -10,7 +10,7 @@ WORKDIR /app RUN apt-get update && apt-get install -y --no-install-recommends curl && rm -rf /var/lib/apt/lists/* -COPY pyproject.toml . +COPY pyproject.toml README.md ./ COPY protoforge/ protoforge/ COPY --from=frontend-builder /app/web/dist /app/static diff --git a/protoforge/main.py b/protoforge/main.py index 894c5bc..0e43951 100644 --- a/protoforge/main.py +++ b/protoforge/main.py @@ -1,8 +1,11 @@ import logging from contextlib import asynccontextmanager +from pathlib import Path from fastapi import FastAPI from fastapi.middleware.cors import CORSMiddleware +from fastapi.responses import FileResponse +from fastapi.staticfiles import StaticFiles from protoforge.api.v1.router import router from protoforge.core.engine import SimulationEngine @@ -184,12 +187,19 @@ def create_app() -> FastAPI: @app.get("/") async def root(): + index = Path("/app/static/index.html") + if index.exists(): + return FileResponse(index) return { "name": "ProtoForge", "version": "0.1.0", "description": "物联网协议仿真与测试平台", } + static_dir = Path("/app/static") + if static_dir.exists(): + app.mount("/assets", StaticFiles(directory=static_dir / "assets"), name="assets") + @app.get("/health") async def health(): return {"status": "ok"} diff --git a/web/src/App.vue b/web/src/App.vue index ae6d315..22eabe9 100644 --- a/web/src/App.vue +++ b/web/src/App.vue @@ -1,4 +1,7 @@ From 550d8e20b98481a5390be716104142d6aec87770 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 20 May 2026 19:57:06 +0800 Subject: [PATCH 17/36] feat(ai): support ai --- ai/predict.py | 97 +++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 97 insertions(+) create mode 100755 ai/predict.py diff --git a/ai/predict.py b/ai/predict.py new file mode 100755 index 0000000..b70f822 --- /dev/null +++ b/ai/predict.py @@ -0,0 +1,97 @@ +# -*- coding: utf-8 -*- + +import requests +import numpy as np +from datetime import datetime, timedelta + +VM_URL = "http://localhost:8428" +DEVICE_ID = "fanuc-cnc" +METRIC = f'feed_rate{{device_id="{DEVICE_ID}"}}' + +def fetch_history(minutes=30): + """从VM拉取历史数据""" + end = datetime.now() + start = end - timedelta(minutes=minutes) + resp = requests.get(f"{VM_URL}/api/v1/query_range", params={ + "query": METRIC, + "start": start.timestamp(), + "end": end.timestamp(), + "step": "1s", + }) + result = resp.json()["data"]["result"] + if not result: + return [], [] + values = result[0]["values"] + ts = [float(v[0]) for v in values] + ys = [float(v[1]) for v in values] + return ts, ys + +def predict_next(ts, ys, horizon=60): + """ + 用FFT检测主频，拟合正弦波，外推未来horizon秒 + 适合周期性信号 + """ + if len(ys) < 60: + return [], [] + + ys = np.array(ys) + n = len(ys) + dt = 1.0 # 1秒采样 + + # FFT找主频 + fft = np.fft.rfft(ys - ys.mean()) + freqs = np.fft.rfftfreq(n, d=dt) + dominant_idx = np.argmax(np.abs(fft[1:])) + 1 + dominant_freq = freqs[dominant_idx] + period = 1.0 / dominant_freq if dominant_freq > 0 else 60 + + # 拟合：y = A*sin(2π/T * t + φ) + offset + from scipy.optimize import curve_fit + t_rel = np.arange(n, dtype=float) + offset = ys.mean() + amplitude = (ys.max() - ys.min()) / 2 + + def sine_model(t, A, T, phi, C): + return A * np.sin(2 * np.pi / T * t + phi) + C + + try: + popt, _ = curve_fit( + sine_model, t_rel, ys, + p0=[amplitude, period, 0, offset], + maxfev=5000 + ) + # 外推 + t_future = np.arange(n, n + horizon, dtype=float) + y_pred = sine_model(t_future, *popt) + ts_future = [ts[-1] + i + 1 for i in range(horizon)] + return ts_future, y_pred.tolist() + except Exception: + # 拟合失败降级为线性 + slope = (ys[-1] - ys[-10]) / 10 + ts_future = [ts[-1] + i + 1 for i in range(horizon)] + y_pred = [ys[-1] + slope * (i + 1) for i in range(horizon)] + return ts_future, y_pred + +def write_predictions(ts_future, y_pred, metric_name="protoforge_feed_rate_predicted"): + """写回VictoriaMetrics""" + lines = [] + for t, y in zip(ts_future, y_pred): + ts_ms = int(t * 1000) + lines.append(f'{metric_name}{{device_id="{DEVICE_ID}"}} {y:.2f} {ts_ms}') + payload = "\n".join(lines) + requests.post(f"{VM_URL}/api/v1/import/prometheus", data=payload) + +def run_once(): + ts, ys = fetch_history(minutes=30) + if len(ys) < 60: + print("数据不足") + return + ts_future, y_pred = predict_next(ts, ys, horizon=120) + write_predictions(ts_future, y_pred) + print(f"写入 {len(y_pred)} 个预测点，预测到 +{len(y_pred)}s") + +if __name__ == "__main__": + import time + while True: + run_once() + time.sleep(30) # 每30秒重新预测一次 From e8f70d09c27d53bff6c8f310b7a2064496632a33 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 20 May 2026 20:29:13 +0800 Subject: [PATCH 18/36] feat(predict_v2): add predict_v2 python file --- ai/predict_v2.py | 206 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 206 insertions(+) create mode 100755 ai/predict_v2.py diff --git a/ai/predict_v2.py b/ai/predict_v2.py new file mode 100755 index 0000000..df5dd97 --- /dev/null +++ b/ai/predict_v2.py @@ -0,0 +1,206 @@ +# -*- coding: utf-8 -*- +""" +ProtoForge 预测服务 v2 +从 VictoriaMetrics 拉取历史数据，用 FFT + 正弦拟合预测未来值，写回 VM。 +预测值时间戳为未来时间，Grafana 中预测线出现在实测线右侧延伸处。 +""" + +import logging +import time +from datetime import datetime, timedelta + +import numpy as np +import requests +from scipy.optimize import curve_fit + +logging.basicConfig( + level=logging.INFO, + format="%(asctime)s [%(levelname)s] %(message)s", +) +logger = logging.getLogger(__name__) + +# ── 配置 ────────────────────────────────────────────────────────────────────── +VM_URL = "http://localhost:8428" + +# 要预测的指标列表，每项：(查询表达式, 写回指标名) +PREDICT_TARGETS = [ + ('feed_rate{device_id="fanuc-cnc"}', "feed_rate_predicted"), + ('spindle_speed{device_id="fanuc-cnc"}', "spindle_speed_predicted"), + ('spindle_current{device_id="fanuc-cnc"}', "spindle_current_predicted"), + ('vibration_x{device_id="fanuc-cnc"}', "vibration_x_predicted"), + ('vibration_y{device_id="fanuc-cnc"}', "vibration_y_predicted"), + ('vibration_z{device_id="fanuc-cnc"}', "vibration_z_predicted"), +] + +HISTORY_MINUTES = 30 # 拉取多少分钟历史数据用于拟合 +HORIZON_SECONDS = 120 # 预测未来多少秒 +POLL_INTERVAL = 30 # 每隔多少秒重新预测一次 +MIN_POINTS = 120 # 至少需要多少个历史点才开始预测 +# ───────────────────────────────────────────────────────────────────────────── + + +def fetch_history(query: str, minutes: int = HISTORY_MINUTES): + """从 VictoriaMetrics 拉取历史时序数据，返回 (timestamps, values)。""" + now = datetime.now() + start = now - timedelta(minutes=minutes) + try: + resp = requests.get( + f"{VM_URL}/api/v1/query_range", + params={ + "query": query, + "start": start.timestamp(), + "end": now.timestamp(), + "step": "1s", + }, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("拉取数据失败 query=%s: %s", query, e) + return [], [] + + result = resp.json().get("data", {}).get("result", []) + if not result: + return [], [] + + values = result[0]["values"] + ts = [float(v[0]) for v in values] + ys = [float(v[1]) for v in values] + return ts, ys + + +def _sine_model(t, A, T, phi, C): + return A * np.sin(2 * np.pi / T * t + phi) + C + + +def predict_next(ts: list, ys: list, horizon: int = HORIZON_SECONDS): + """ + 用 FFT 检测主频，拟合正弦波，外推未来 horizon 秒。 + 返回 (future_timestamps, predicted_values)，时间戳均在最后一个真实点之后。 + 降级策略：拟合失败时用最近 10 点线性外推。 + """ + ys_arr = np.array(ys) + n = len(ys_arr) + + # ── FFT 找主频 ──────────────────────────────────────────────────────────── + fft_vals = np.fft.rfft(ys_arr - ys_arr.mean()) + freqs = np.fft.rfftfreq(n, d=1.0) # d=1 表示 1 秒采样间隔 + # 跳过直流分量（index 0） + dominant_idx = int(np.argmax(np.abs(fft_vals[1:]))) + 1 + dominant_freq = freqs[dominant_idx] + period = 1.0 / dominant_freq if dominant_freq > 0 else 60.0 + period = float(np.clip(period, 5.0, 3600.0)) # 限制在合理范围 + + # ── 正弦拟合 ────────────────────────────────────────────────────────────── + t_rel = np.arange(n, dtype=float) + amplitude = (ys_arr.max() - ys_arr.min()) / 2.0 + offset = float(ys_arr.mean()) + + # 最后一个真实数据点的 Unix 时间戳（秒） + last_ts = ts[-1] + + try: + popt, _ = curve_fit( + _sine_model, + t_rel, + ys_arr, + p0=[amplitude, period, 0.0, offset], + bounds=( + [0, 5.0, -np.pi, ys_arr.min()], + [np.inf, 3600.0, np.pi, ys_arr.max()], + ), + maxfev=8000, + ) + t_future = np.arange(n, n + horizon, dtype=float) + y_pred = _sine_model(t_future, *popt) + # 裁剪到历史数据值域，避免外推飞出合理范围 + y_pred = np.clip(y_pred, ys_arr.min() * 0.5, ys_arr.max() * 1.5) + + # 未来时间戳：last_ts + 1s, +2s, ..., +horizon s + ts_future = [last_ts + i + 1 for i in range(horizon)] + logger.debug("正弦拟合成功 period=%.1fs amplitude=%.2f", popt[1], popt[0]) + return ts_future, y_pred.tolist() + + except Exception as e: + logger.warning("正弦拟合失败，降级为线性外推: %s", e) + tail = min(10, n) + slope = (ys_arr[-1] - ys_arr[-tail]) / tail + ts_future = [last_ts + i + 1 for i in range(horizon)] + y_pred = [float(ys_arr[-1] + slope * (i + 1)) for i in range(horizon)] + return ts_future, y_pred + + +def write_predictions(ts_future: list, y_pred: list, metric_name: str, extra_labels: dict = None): + """ + 将预测值以 Prometheus exposition 格式写入 VictoriaMetrics。 + 时间戳为毫秒级 Unix 时间戳，对应未来时间点。 + """ + label_str = "" + if extra_labels: + parts = [f'{k}="{v}"' for k, v in extra_labels.items()] + label_str = "{" + ",".join(parts) + "}" + + lines = [] + for t, y in zip(ts_future, y_pred): + ts_ms = int(t * 1000) + lines.append(f"{metric_name}{label_str} {y:.4f} {ts_ms}") + + payload = "\n".join(lines) + try: + resp = requests.post( + f"{VM_URL}/api/v1/import/prometheus", + data=payload, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("写入预测数据失败 metric=%s: %s", metric_name, e) + + +def _parse_labels(query: str) -> dict: + """从查询表达式中解析标签，如 feed_rate{device_id="fanuc-cnc"} → {"device_id": "fanuc-cnc"}""" + labels = {} + if "{" not in query: + return labels + label_part = query[query.index("{") + 1: query.index("}")] + for item in label_part.split(","): + if "=" in item: + k, v = item.split("=", 1) + labels[k.strip()] = v.strip().strip('"') + return labels + + +def run_once(): + now_str = datetime.now().strftime("%H:%M:%S") + for query, pred_metric in PREDICT_TARGETS: + ts, ys = fetch_history(query) + if len(ys) < MIN_POINTS: + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + continue + + ts_future, y_pred = predict_next(ts, ys, horizon=HORIZON_SECONDS) + if not ts_future: + continue + + extra_labels = _parse_labels(query) + write_predictions(ts_future, y_pred, pred_metric, extra_labels) + + future_time = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + logger.info( + "[%s] %-40s → %-35s 写入 %d 点，预测至 %s", + now_str, query, pred_metric, len(y_pred), future_time, + ) + + +def main(): + logger.info( + "预测服务启动 VM=%s 预测窗口=%ds 轮询间隔=%ds", + VM_URL, HORIZON_SECONDS, POLL_INTERVAL, + ) + while True: + run_once() + time.sleep(POLL_INTERVAL) + + +if __name__ == "__main__": + main() From 57df20284645347aba7de4e2399640e05b6d0b3c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 20 May 2026 21:13:52 +0800 Subject: [PATCH 19/36] fix --- ai/predict_v2.py | 536 +++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 450 insertions(+), 86 deletions(-) diff --git a/ai/predict_v2.py b/ai/predict_v2.py index df5dd97..bc425c8 100755 --- a/ai/predict_v2.py +++ b/ai/predict_v2.py @@ -1,56 +1,93 @@ # -*- coding: utf-8 -*- """ -ProtoForge 预测服务 v2 -从 VictoriaMetrics 拉取历史数据，用 FFT + 正弦拟合预测未来值，写回 VM。 -预测值时间戳为未来时间，Grafana 中预测线出现在实测线右侧延伸处。 +ProtoForge 预测服务 v3 + +修复点： +1. 解决 HORIZON_SECONDS > POLL_INTERVAL 时，多轮预测窗口重叠导致 Grafana 出现毛刺/竖线问题。 +2. 每轮写入新预测前，删除同一个预测 metric 的旧预测序列，只保留最新一轮预测。 +3. 预测时间戳按整秒写入，避免毫秒时间戳和 Grafana step 不对齐。 +4. 拟合使用真实 timestamp 相对时间，不再假设历史数据严格 1 秒等间隔。 +5. 对历史数据做排序、去重、NaN/Inf 清洗。 """ import logging +import math +import re import time from datetime import datetime, timedelta +from typing import Dict, List, Tuple import numpy as np import requests from scipy.optimize import curve_fit + logging.basicConfig( level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s", ) + logger = logging.getLogger(__name__) + # ── 配置 ────────────────────────────────────────────────────────────────────── + VM_URL = "http://localhost:8428" -# 要预测的指标列表，每项：(查询表达式, 写回指标名) PREDICT_TARGETS = [ - ('feed_rate{device_id="fanuc-cnc"}', "feed_rate_predicted"), - ('spindle_speed{device_id="fanuc-cnc"}', "spindle_speed_predicted"), - ('spindle_current{device_id="fanuc-cnc"}', "spindle_current_predicted"), - ('vibration_x{device_id="fanuc-cnc"}', "vibration_x_predicted"), - ('vibration_y{device_id="fanuc-cnc"}', "vibration_y_predicted"), - ('vibration_z{device_id="fanuc-cnc"}', "vibration_z_predicted"), + ('feed_rate{device_id="fanuc-cnc"}', "feed_rate_predicted"), + ('spindle_speed{device_id="fanuc-cnc"}', "spindle_speed_predicted"), + ('spindle_current{device_id="fanuc-cnc"}', "spindle_current_predicted"), + ('vibration_x{device_id="fanuc-cnc"}', "vibration_x_predicted"), + ('vibration_y{device_id="fanuc-cnc"}', "vibration_y_predicted"), + ('vibration_z{device_id="fanuc-cnc"}', "vibration_z_predicted"), ] -HISTORY_MINUTES = 30 # 拉取多少分钟历史数据用于拟合 -HORIZON_SECONDS = 120 # 预测未来多少秒 -POLL_INTERVAL = 30 # 每隔多少秒重新预测一次 -MIN_POINTS = 120 # 至少需要多少个历史点才开始预测 +HISTORY_MINUTES = 30 +HORIZON_SECONDS = 120 +POLL_INTERVAL = 30 +MIN_POINTS = 120 +QUERY_STEP = "1s" + +# 关键修复：每轮写入前删除旧预测，避免 120s 预测窗口和 30s 轮询周期重叠 +CLEAR_OLD_PREDICTIONS = True + +# 如果删除旧预测失败，是否跳过本轮写入。 +# 建议 True，避免继续叠加脏数据。 +SKIP_WRITE_IF_CLEAR_FAILED = True + +# 给新预测数据加一个稳定标签，方便 Grafana 查询过滤。 +# Grafana 可以查询：feed_rate_predicted{device_id="fanuc-cnc",forecast="latest"} +EXTRA_PREDICT_LABELS = { + "forecast": "latest", + "source": "protoforge", +} + +# 正弦周期限制 +MIN_PERIOD_SECONDS = 5.0 +MAX_PERIOD_SECONDS = 3600.0 + # ───────────────────────────────────────────────────────────────────────────── -def fetch_history(query: str, minutes: int = HISTORY_MINUTES): - """从 VictoriaMetrics 拉取历史时序数据，返回 (timestamps, values)。""" +def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: + """ + 从 VictoriaMetrics 拉取历史时序数据。 + 返回： + timestamps: Unix 秒级时间戳 + values: float 数值 + """ now = datetime.now() start = now - timedelta(minutes=minutes) + try: resp = requests.get( f"{VM_URL}/api/v1/query_range", params={ "query": query, "start": start.timestamp(), - "end": now.timestamp(), - "step": "1s", + "end": now.timestamp(), + "step": QUERY_STEP, }, timeout=10, ) @@ -59,148 +96,475 @@ def fetch_history(query: str, minutes: int = HISTORY_MINUTES): logger.error("拉取数据失败 query=%s: %s", query, e) return [], [] - result = resp.json().get("data", {}).get("result", []) + try: + result = resp.json().get("data", {}).get("result", []) + except Exception as e: + logger.error("解析 VM 返回失败 query=%s: %s", query, e) + return [], [] + if not result: return [], [] - values = result[0]["values"] - ts = [float(v[0]) for v in values] - ys = [float(v[1]) for v in values] + values = result[0].get("values", []) + if not values: + return [], [] + + ts = [] + ys = [] + + for item in values: + if len(item) < 2: + continue + + try: + t = float(item[0]) + y = float(item[1]) + except Exception: + continue + + if not math.isfinite(t) or not math.isfinite(y): + continue + + ts.append(t) + ys.append(y) + return ts, ys -def _sine_model(t, A, T, phi, C): - return A * np.sin(2 * np.pi / T * t + phi) + C +def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: + """ + 清洗历史数据： + 1. 转换为整秒时间戳 + 2. 排序 + 3. 同一秒多个值时保留最后一个 + 4. 插值补齐中间缺失秒 + """ + if not ts or not ys or len(ts) != len(ys): + return np.array([]), np.array([]) + + data = {} + + for t, y in zip(ts, ys): + try: + sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(sec) or not math.isfinite(val): + continue + + data[sec] = val + + if not data: + return np.array([]), np.array([]) + + sorted_items = sorted(data.items(), key=lambda x: x[0]) + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) + ys_clean = np.array([x[1] for x in sorted_items], dtype=float) -def predict_next(ts: list, ys: list, horizon: int = HORIZON_SECONDS): + if len(ts_clean) < 2: + return ts_clean, ys_clean + + start_sec = int(ts_clean[0]) + end_sec = int(ts_clean[-1]) + + if end_sec <= start_sec: + return ts_clean, ys_clean + + # 统一为 1 秒网格，减少 query_range 缺点、抖动、缺失点对 FFT 的影响 + ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) + ys_grid = np.interp(ts_grid, ts_clean, ys_clean) + + return ts_grid, ys_grid + + +def _sine_model(t: np.ndarray, A: float, T: float, phi: float, C: float) -> np.ndarray: + return A * np.sin(2.0 * np.pi / T * t + phi) + C + + +def estimate_period_by_fft(ys_arr: np.ndarray) -> float: """ - 用 FFT 检测主频，拟合正弦波，外推未来 horizon 秒。 - 返回 (future_timestamps, predicted_values)，时间戳均在最后一个真实点之后。 - 降级策略：拟合失败时用最近 10 点线性外推。 + 使用 FFT 估算主周期。 + ys_arr 默认是 1 秒间隔。 """ - ys_arr = np.array(ys) n = len(ys_arr) - # ── FFT 找主频 ──────────────────────────────────────────────────────────── - fft_vals = np.fft.rfft(ys_arr - ys_arr.mean()) - freqs = np.fft.rfftfreq(n, d=1.0) # d=1 表示 1 秒采样间隔 - # 跳过直流分量（index 0） - dominant_idx = int(np.argmax(np.abs(fft_vals[1:]))) + 1 - dominant_freq = freqs[dominant_idx] - period = 1.0 / dominant_freq if dominant_freq > 0 else 60.0 - period = float(np.clip(period, 5.0, 3600.0)) # 限制在合理范围 + if n < 4: + return 60.0 + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return 60.0 + + fft_vals = np.fft.rfft(centered) + freqs = np.fft.rfftfreq(n, d=1.0) + + if len(freqs) <= 1: + return 60.0 + + # 跳过直流分量 index 0 + power = np.abs(fft_vals[1:]) + if len(power) == 0 or np.max(power) <= 0: + return 60.0 + + dominant_idx = int(np.argmax(power)) + 1 + dominant_freq = float(freqs[dominant_idx]) + + if dominant_freq <= 0: + return 60.0 + + period = 1.0 / dominant_freq + period = float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + return period - # ── 正弦拟合 ────────────────────────────────────────────────────────────── - t_rel = np.arange(n, dtype=float) - amplitude = (ys_arr.max() - ys_arr.min()) / 2.0 - offset = float(ys_arr.mean()) - # 最后一个真实数据点的 Unix 时间戳（秒） - last_ts = ts[-1] +def predict_next( + ts: List[float], + ys: List[float], + horizon: int = HORIZON_SECONDS, + start_from_now: bool = True, +) -> Tuple[List[float], List[float]]: + """ + 用 FFT 检测主频，拟合正弦波，外推未来 horizon 秒。 + 返回： + future_timestamps: 未来整秒时间戳 + predicted_values: 预测值 + """ + ts_grid, ys_grid = normalize_history(ts, ys) + + if len(ys_grid) < MIN_POINTS: + return [], [] + + n = len(ys_grid) + + y_min = float(np.min(ys_grid)) + y_max = float(np.max(ys_grid)) + y_mean = float(np.mean(ys_grid)) + y_range = y_max - y_min + + # 数据几乎不波动时，直接使用最后一个值保持 + if y_range <= 1e-9: + base_ts = int(time.time()) if start_from_now else int(ts_grid[-1]) + base_ts = max(base_ts, int(ts_grid[-1])) + + ts_future = [base_ts + i + 1 for i in range(horizon)] + y_pred = [float(ys_grid[-1])] * horizon + return ts_future, y_pred + + period = estimate_period_by_fft(ys_grid) + + # 用真实时间戳做相对时间，而不是 np.arange(n) + t_fit = ts_grid - ts_grid[0] + + amplitude = y_range / 2.0 + offset = y_mean + + # 预测起点统一对齐到整秒 + if start_from_now: + base_ts = int(time.time()) + else: + base_ts = int(ts_grid[-1]) + + # 避免因为 VM 查询延迟导致预测点落在最后一个真实点之前 + base_ts = max(base_ts, int(ts_grid[-1])) + + ts_future_arr = np.arange(base_ts + 1, base_ts + 1 + horizon, 1, dtype=float) + t_future = ts_future_arr - ts_grid[0] try: popt, _ = curve_fit( _sine_model, - t_rel, - ys_arr, + t_fit, + ys_grid, p0=[amplitude, period, 0.0, offset], bounds=( - [0, 5.0, -np.pi, ys_arr.min()], - [np.inf, 3600.0, np.pi, ys_arr.max()], + [0.0, MIN_PERIOD_SECONDS, -2.0 * np.pi, y_min - y_range], + [np.inf, MAX_PERIOD_SECONDS, 2.0 * np.pi, y_max + y_range], ), - maxfev=8000, + maxfev=12000, + ) + + y_pred_arr = _sine_model(t_future, *popt) + + # 裁剪到合理范围，避免拟合异常时飞出去 + margin = y_range * 0.2 + lower = y_min - margin + upper = y_max + margin + y_pred_arr = np.clip(y_pred_arr, lower, upper) + + if not np.all(np.isfinite(y_pred_arr)): + raise ValueError("预测结果包含 NaN/Inf") + + logger.debug( + "正弦拟合成功 period=%.2fs amplitude=%.4f offset=%.4f", + popt[1], + popt[0], + popt[3], ) - t_future = np.arange(n, n + horizon, dtype=float) - y_pred = _sine_model(t_future, *popt) - # 裁剪到历史数据值域，避免外推飞出合理范围 - y_pred = np.clip(y_pred, ys_arr.min() * 0.5, ys_arr.max() * 1.5) - # 未来时间戳：last_ts + 1s, +2s, ..., +horizon s - ts_future = [last_ts + i + 1 for i in range(horizon)] - logger.debug("正弦拟合成功 period=%.1fs amplitude=%.2f", popt[1], popt[0]) - return ts_future, y_pred.tolist() + return ts_future_arr.tolist(), y_pred_arr.astype(float).tolist() except Exception as e: - logger.warning("正弦拟合失败，降级为线性外推: %s", e) + logger.warning("正弦拟合失败，降级为最近值平滑外推: %s", e) + + # 降级策略：用最近 10 个点的均值保持，避免线性外推越走越偏 tail = min(10, n) - slope = (ys_arr[-1] - ys_arr[-tail]) / tail - ts_future = [last_ts + i + 1 for i in range(horizon)] - y_pred = [float(ys_arr[-1] + slope * (i + 1)) for i in range(horizon)] + last_value = float(np.mean(ys_grid[-tail:])) + + ts_future = ts_future_arr.tolist() + y_pred = [last_value] * horizon + return ts_future, y_pred -def write_predictions(ts_future: list, y_pred: list, metric_name: str, extra_labels: dict = None): +def prom_escape_label_value(value: str) -> str: + """ + Prometheus exposition label value 转义。 + """ + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) + + +def build_selector(metric_name: str, labels: Dict[str, str]) -> str: + """ + 构造 PromQL selector，用于 delete_series。 + + 示例： + feed_rate_predicted{device_id="fanuc-cnc"} + """ + if not labels: + return metric_name + + parts = [] + for k in sorted(labels.keys()): + v = prom_escape_label_value(labels[k]) + parts.append(f'{k}="{v}"') + + return f'{metric_name}' + "{" + ",".join(parts) + "}" + + +def delete_old_predictions(metric_name: str, base_labels: Dict[str, str]) -> bool: + """ + 删除旧预测序列，避免多轮预测窗口重叠。 + + 注意： + 这里故意只用 base_labels，比如 device_id。 + 不带 forecast/source 标签，是为了兼容旧版本脚本写入的无 forecast 标签数据。 + """ + selector = build_selector(metric_name, base_labels) + + try: + resp = requests.post( + f"{VM_URL}/api/v1/admin/tsdb/delete_series", + params=[("match[]", selector)], + timeout=10, + ) + + if resp.status_code not in (200, 204): + logger.error( + "删除旧预测数据失败 metric=%s selector=%s status=%s body=%s", + metric_name, + selector, + resp.status_code, + resp.text[:500], + ) + return False + + logger.debug("已删除旧预测数据 selector=%s", selector) + return True + + except requests.RequestException as e: + logger.error("删除旧预测数据异常 metric=%s selector=%s: %s", metric_name, selector, e) + return False + + +def write_predictions( + ts_future: List[float], + y_pred: List[float], + metric_name: str, + labels: Dict[str, str] = None, +) -> bool: """ 将预测值以 Prometheus exposition 格式写入 VictoriaMetrics。 - 时间戳为毫秒级 Unix 时间戳，对应未来时间点。 + 时间戳为毫秒级 Unix timestamp。 """ + if labels is None: + labels = {} + + if not ts_future or not y_pred or len(ts_future) != len(y_pred): + logger.warning("预测数据为空或长度不一致 metric=%s", metric_name) + return False + label_str = "" - if extra_labels: - parts = [f'{k}="{v}"' for k, v in extra_labels.items()] + if labels: + parts = [] + for k in sorted(labels.keys()): + v = prom_escape_label_value(labels[k]) + parts.append(f'{k}="{v}"') label_str = "{" + ",".join(parts) + "}" lines = [] + for t, y in zip(ts_future, y_pred): - ts_ms = int(t * 1000) - lines.append(f"{metric_name}{label_str} {y:.4f} {ts_ms}") + try: + ts_sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(ts_sec) or not math.isfinite(val): + continue + + ts_ms = ts_sec * 1000 + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_ms}") + + if not lines: + logger.warning("没有可写入的预测点 metric=%s", metric_name) + return False + + payload = "\n".join(lines) + "\n" - payload = "\n".join(lines) try: resp = requests.post( f"{VM_URL}/api/v1/import/prometheus", - data=payload, + data=payload.encode("utf-8"), + headers={ + "Content-Type": "text/plain; version=0.0.4; charset=utf-8", + }, timeout=10, ) resp.raise_for_status() + return True + except requests.RequestException as e: logger.error("写入预测数据失败 metric=%s: %s", metric_name, e) + return False + + +_LABEL_PATTERN = re.compile(r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*') -def _parse_labels(query: str) -> dict: - """从查询表达式中解析标签，如 feed_rate{device_id="fanuc-cnc"} → {"device_id": "fanuc-cnc"}""" +def _parse_labels(query: str) -> Dict[str, str]: + """ + 从查询表达式中解析标签。 + + 示例： + feed_rate{device_id="fanuc-cnc"} -> {"device_id": "fanuc-cnc"} + """ labels = {} - if "{" not in query: + + if "{" not in query or "}" not in query: return labels - label_part = query[query.index("{") + 1: query.index("}")] - for item in label_part.split(","): - if "=" in item: - k, v = item.split("=", 1) - labels[k.strip()] = v.strip().strip('"') + + try: + label_part = query[query.index("{") + 1: query.rindex("}")] + except Exception: + return labels + + for match in _LABEL_PATTERN.finditer(label_part): + key = match.group(1) + value = match.group(2) + value = value.replace('\\"', '"').replace("\\n", "\n").replace("\\\\", "\\") + labels[key] = value + return labels +def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: + result = {} + + for d in dicts: + if not d: + continue + result.update(d) + + return result + + def run_once(): now_str = datetime.now().strftime("%H:%M:%S") + for query, pred_metric in PREDICT_TARGETS: ts, ys = fetch_history(query) + if len(ys) < MIN_POINTS: logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) continue - ts_future, y_pred = predict_next(ts, ys, horizon=HORIZON_SECONDS) - if not ts_future: + ts_future, y_pred = predict_next( + ts, + ys, + horizon=HORIZON_SECONDS, + start_from_now=True, + ) + + if not ts_future or not y_pred: + logger.warning("[%s] %s 预测结果为空，跳过", now_str, query) + continue + + base_labels = _parse_labels(query) + + # 先删除旧预测，再写入新预测。 + # 删除条件只带 base_labels，兼容老版本无 forecast/source 标签的脏数据。 + if CLEAR_OLD_PREDICTIONS: + clear_ok = delete_old_predictions(pred_metric, base_labels) + + if not clear_ok and SKIP_WRITE_IF_CLEAR_FAILED: + logger.error( + "[%s] %s 删除旧预测失败，为避免继续制造重叠数据，本轮跳过写入", + now_str, + pred_metric, + ) + continue + + write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + + ok = write_predictions( + ts_future=ts_future, + y_pred=y_pred, + metric_name=pred_metric, + labels=write_labels, + ) + + if not ok: continue - extra_labels = _parse_labels(query) - write_predictions(ts_future, y_pred, pred_metric, extra_labels) + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") + future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") - future_time = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") logger.info( - "[%s] %-40s → %-35s 写入 %d 点，预测至 %s", - now_str, query, pred_metric, len(y_pred), future_time, + "[%s] %-40s → %-35s 写入 %d 点，预测区间 %s ~ %s", + now_str, + query, + pred_metric, + len(y_pred), + future_start, + future_end, ) def main(): logger.info( - "预测服务启动 VM=%s 预测窗口=%ds 轮询间隔=%ds", - VM_URL, HORIZON_SECONDS, POLL_INTERVAL, + "预测服务启动 VM=%s 历史窗口=%dmin 预测窗口=%ds 轮询间隔=%ds 清理旧预测=%s", + VM_URL, + HISTORY_MINUTES, + HORIZON_SECONDS, + POLL_INTERVAL, + CLEAR_OLD_PREDICTIONS, ) + while True: run_once() time.sleep(POLL_INTERVAL) if __name__ == "__main__": - main() + main() \ No newline at end of file From 88aec295671ca112fd422a28acd4d76d43a82f20 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 20 May 2026 21:21:11 +0800 Subject: [PATCH 20/36] fix --- ai/predict_v2.py | 263 +++++++++++++++++++++++------------------------ 1 file changed, 128 insertions(+), 135 deletions(-) diff --git a/ai/predict_v2.py b/ai/predict_v2.py index bc425c8..f631e12 100755 --- a/ai/predict_v2.py +++ b/ai/predict_v2.py @@ -1,13 +1,14 @@ # -*- coding: utf-8 -*- """ -ProtoForge 预测服务 v3 +ProtoForge 预测服务 v4 修复点： -1. 解决 HORIZON_SECONDS > POLL_INTERVAL 时，多轮预测窗口重叠导致 Grafana 出现毛刺/竖线问题。 -2. 每轮写入新预测前，删除同一个预测 metric 的旧预测序列，只保留最新一轮预测。 -3. 预测时间戳按整秒写入，避免毫秒时间戳和 Grafana step 不对齐。 -4. 拟合使用真实 timestamp 相对时间，不再假设历史数据严格 1 秒等间隔。 -5. 对历史数据做排序、去重、NaN/Inf 清洗。 +1. 不再使用 VictoriaMetrics delete_series，避免预测历史被整条删除。 +2. 不再每 30 秒写未来 120 秒，避免多轮预测窗口重叠导致 Grafana 出现竖线/毛刺。 +3. 每轮只写未来 min(HORIZON_SECONDS, POLL_INTERVAL) 秒的数据。 +4. 使用 forecast="rolling_v2" 新标签，避免和上一版 forecast="latest" 的旧预测数据混在一起。 +5. 使用真实 timestamp 做拟合，不假设采样严格等间隔。 +6. 拟合失败时不再简单写平直线，而是尽量重复最近一个周期的波形。 """ import logging @@ -44,29 +45,35 @@ ] HISTORY_MINUTES = 30 + +# 理论预测窗口 HORIZON_SECONDS = 120 + +# 轮询间隔 POLL_INTERVAL = 30 + +# 实际写入窗口。 +# 关键点：实际写入窗口不要大于轮询间隔，否则不同批次预测会重叠。 +WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) + MIN_POINTS = 120 QUERY_STEP = "1s" -# 关键修复：每轮写入前删除旧预测，避免 120s 预测窗口和 30s 轮询周期重叠 -CLEAR_OLD_PREDICTIONS = True - -# 如果删除旧预测失败，是否跳过本轮写入。 -# 建议 True，避免继续叠加脏数据。 -SKIP_WRITE_IF_CLEAR_FAILED = True +# 不要再清理旧预测，否则历史预测会被整条删除。 +CLEAR_OLD_PREDICTIONS = False -# 给新预测数据加一个稳定标签，方便 Grafana 查询过滤。 -# Grafana 可以查询：feed_rate_predicted{device_id="fanuc-cnc",forecast="latest"} +# 使用新标签，避免和上一版 forecast="latest" 数据混在一起。 EXTRA_PREDICT_LABELS = { - "forecast": "latest", + "forecast": "rolling_v2", "source": "protoforge", } -# 正弦周期限制 MIN_PERIOD_SECONDS = 5.0 MAX_PERIOD_SECONDS = 3600.0 +# 进程内记录每条预测序列上次写到哪里，避免本进程运行期间重复写同一时间段 +LAST_WRITTEN_UNTIL: Dict[str, int] = {} + # ───────────────────────────────────────────────────────────────────────────── @@ -134,10 +141,10 @@ def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[floa def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: """ 清洗历史数据： - 1. 转换为整秒时间戳 + 1. 时间戳转为整秒 2. 排序 3. 同一秒多个值时保留最后一个 - 4. 插值补齐中间缺失秒 + 4. 插值补齐缺失秒 """ if not ts or not ys or len(ts) != len(ys): return np.array([]), np.array([]) @@ -173,7 +180,6 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. if end_sec <= start_sec: return ts_clean, ys_clean - # 统一为 1 秒网格，减少 query_range 缺点、抖动、缺失点对 FFT 的影响 ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) ys_grid = np.interp(ts_grid, ts_clean, ys_clean) @@ -187,7 +193,7 @@ def _sine_model(t: np.ndarray, A: float, T: float, phi: float, C: float) -> np.n def estimate_period_by_fft(ys_arr: np.ndarray) -> float: """ 使用 FFT 估算主周期。 - ys_arr 默认是 1 秒间隔。 + ys_arr 默认已经是 1 秒间隔。 """ n = len(ys_arr) @@ -205,8 +211,8 @@ def estimate_period_by_fft(ys_arr: np.ndarray) -> float: if len(freqs) <= 1: return 60.0 - # 跳过直流分量 index 0 power = np.abs(fft_vals[1:]) + if len(power) == 0 or np.max(power) <= 0: return 60.0 @@ -222,59 +228,84 @@ def estimate_period_by_fft(ys_arr: np.ndarray) -> float: return period +def repeat_last_period( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + ts_future_arr: np.ndarray, + period_seconds: float, +) -> np.ndarray: + """ + 拟合失败时的降级策略： + 不直接写平直线，而是把未来时间映射回最近一个周期的历史波形。 + """ + if len(ts_grid) < 2: + return np.full_like(ts_future_arr, float(ys_grid[-1]), dtype=float) + + period = max(int(round(period_seconds)), 1) + + y_pred = [] + + hist_start = float(ts_grid[0]) + hist_end = float(ts_grid[-1]) + + for future_ts in ts_future_arr: + mapped_ts = float(future_ts) + + while mapped_ts > hist_end: + mapped_ts -= period + + while mapped_ts < hist_start: + mapped_ts += period + + val = float(np.interp(mapped_ts, ts_grid, ys_grid)) + y_pred.append(val) + + return np.array(y_pred, dtype=float) + + def predict_next( ts: List[float], ys: List[float], - horizon: int = HORIZON_SECONDS, - start_from_now: bool = True, + horizon: int, + base_ts: int, ) -> Tuple[List[float], List[float]]: """ 用 FFT 检测主频，拟合正弦波，外推未来 horizon 秒。 - 返回： - future_timestamps: 未来整秒时间戳 - predicted_values: 预测值 + + base_ts: + 从 base_ts + 1 开始写预测。 """ ts_grid, ys_grid = normalize_history(ts, ys) if len(ys_grid) < MIN_POINTS: return [], [] - n = len(ys_grid) - y_min = float(np.min(ys_grid)) y_max = float(np.max(ys_grid)) y_mean = float(np.mean(ys_grid)) y_range = y_max - y_min - # 数据几乎不波动时，直接使用最后一个值保持 - if y_range <= 1e-9: - base_ts = int(time.time()) if start_from_now else int(ts_grid[-1]) - base_ts = max(base_ts, int(ts_grid[-1])) + base_ts = max(int(base_ts), int(ts_grid[-1])) + + ts_future_arr = np.arange( + base_ts + 1, + base_ts + 1 + horizon, + 1, + dtype=float, + ) - ts_future = [base_ts + i + 1 for i in range(horizon)] - y_pred = [float(ys_grid[-1])] * horizon - return ts_future, y_pred + if y_range <= 1e-9: + y_pred_arr = np.full_like(ts_future_arr, float(ys_grid[-1]), dtype=float) + return ts_future_arr.tolist(), y_pred_arr.tolist() period = estimate_period_by_fft(ys_grid) - # 用真实时间戳做相对时间，而不是 np.arange(n) t_fit = ts_grid - ts_grid[0] + t_future = ts_future_arr - ts_grid[0] amplitude = y_range / 2.0 offset = y_mean - # 预测起点统一对齐到整秒 - if start_from_now: - base_ts = int(time.time()) - else: - base_ts = int(ts_grid[-1]) - - # 避免因为 VM 查询延迟导致预测点落在最后一个真实点之前 - base_ts = max(base_ts, int(ts_grid[-1])) - - ts_future_arr = np.arange(base_ts + 1, base_ts + 1 + horizon, 1, dtype=float) - t_future = ts_future_arr - ts_grid[0] - try: popt, _ = curve_fit( _sine_model, @@ -290,7 +321,6 @@ def predict_next( y_pred_arr = _sine_model(t_future, *popt) - # 裁剪到合理范围，避免拟合异常时飞出去 margin = y_range * 0.2 lower = y_min - margin upper = y_max + margin @@ -309,16 +339,21 @@ def predict_next( return ts_future_arr.tolist(), y_pred_arr.astype(float).tolist() except Exception as e: - logger.warning("正弦拟合失败，降级为最近值平滑外推: %s", e) + logger.warning("正弦拟合失败，降级为最近周期波形复制: %s", e) - # 降级策略：用最近 10 个点的均值保持，避免线性外推越走越偏 - tail = min(10, n) - last_value = float(np.mean(ys_grid[-tail:])) + y_pred_arr = repeat_last_period( + ts_grid=ts_grid, + ys_grid=ys_grid, + ts_future_arr=ts_future_arr, + period_seconds=period, + ) - ts_future = ts_future_arr.tolist() - y_pred = [last_value] * horizon + margin = y_range * 0.2 + lower = y_min - margin + upper = y_max + margin + y_pred_arr = np.clip(y_pred_arr, lower, upper) - return ts_future, y_pred + return ts_future_arr.tolist(), y_pred_arr.astype(float).tolist() def prom_escape_label_value(value: str) -> str: @@ -333,83 +368,34 @@ def prom_escape_label_value(value: str) -> str: ) -def build_selector(metric_name: str, labels: Dict[str, str]) -> str: - """ - 构造 PromQL selector，用于 delete_series。 - - 示例： - feed_rate_predicted{device_id="fanuc-cnc"} - """ +def labels_to_str(labels: Dict[str, str]) -> str: if not labels: - return metric_name + return "" parts = [] + for k in sorted(labels.keys()): v = prom_escape_label_value(labels[k]) parts.append(f'{k}="{v}"') - return f'{metric_name}' + "{" + ",".join(parts) + "}" - - -def delete_old_predictions(metric_name: str, base_labels: Dict[str, str]) -> bool: - """ - 删除旧预测序列，避免多轮预测窗口重叠。 - - 注意： - 这里故意只用 base_labels，比如 device_id。 - 不带 forecast/source 标签，是为了兼容旧版本脚本写入的无 forecast 标签数据。 - """ - selector = build_selector(metric_name, base_labels) - - try: - resp = requests.post( - f"{VM_URL}/api/v1/admin/tsdb/delete_series", - params=[("match[]", selector)], - timeout=10, - ) - - if resp.status_code not in (200, 204): - logger.error( - "删除旧预测数据失败 metric=%s selector=%s status=%s body=%s", - metric_name, - selector, - resp.status_code, - resp.text[:500], - ) - return False - - logger.debug("已删除旧预测数据 selector=%s", selector) - return True - - except requests.RequestException as e: - logger.error("删除旧预测数据异常 metric=%s selector=%s: %s", metric_name, selector, e) - return False + return "{" + ",".join(parts) + "}" def write_predictions( ts_future: List[float], y_pred: List[float], metric_name: str, - labels: Dict[str, str] = None, + labels: Dict[str, str], ) -> bool: """ 将预测值以 Prometheus exposition 格式写入 VictoriaMetrics。 时间戳为毫秒级 Unix timestamp。 """ - if labels is None: - labels = {} - if not ts_future or not y_pred or len(ts_future) != len(y_pred): logger.warning("预测数据为空或长度不一致 metric=%s", metric_name) return False - label_str = "" - if labels: - parts = [] - for k in sorted(labels.keys()): - v = prom_escape_label_value(labels[k]) - parts.append(f'{k}="{v}"') - label_str = "{" + ",".join(parts) + "}" + label_str = labels_to_str(labels) lines = [] @@ -449,7 +435,9 @@ def write_predictions( return False -_LABEL_PATTERN = re.compile(r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*') +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) def _parse_labels(query: str) -> Dict[str, str]: @@ -489,6 +477,13 @@ def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: return result +def series_key(metric_name: str, labels: Dict[str, str]) -> str: + """ + 构造进程内唯一 key，用于记录上次写到哪个时间点。 + """ + return metric_name + labels_to_str(labels) + + def run_once(): now_str = datetime.now().strftime("%H:%M:%S") @@ -499,34 +494,28 @@ def run_once(): logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) continue + base_labels = _parse_labels(query) + write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + + key = series_key(pred_metric, write_labels) + + now_sec = int(time.time()) + last_until = LAST_WRITTEN_UNTIL.get(key, 0) + + # 防止同一进程内重复写入已经预测过的时间段 + base_ts = max(now_sec, last_until) + ts_future, y_pred = predict_next( - ts, - ys, - horizon=HORIZON_SECONDS, - start_from_now=True, + ts=ts, + ys=ys, + horizon=WRITE_HORIZON_SECONDS, + base_ts=base_ts, ) if not ts_future or not y_pred: logger.warning("[%s] %s 预测结果为空，跳过", now_str, query) continue - base_labels = _parse_labels(query) - - # 先删除旧预测，再写入新预测。 - # 删除条件只带 base_labels，兼容老版本无 forecast/source 标签的脏数据。 - if CLEAR_OLD_PREDICTIONS: - clear_ok = delete_old_predictions(pred_metric, base_labels) - - if not clear_ok and SKIP_WRITE_IF_CLEAR_FAILED: - logger.error( - "[%s] %s 删除旧预测失败，为避免继续制造重叠数据，本轮跳过写入", - now_str, - pred_metric, - ) - continue - - write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) - ok = write_predictions( ts_future=ts_future, y_pred=y_pred, @@ -537,26 +526,30 @@ def run_once(): if not ok: continue + LAST_WRITTEN_UNTIL[key] = int(max(ts_future)) + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") logger.info( - "[%s] %-40s → %-35s 写入 %d 点，预测区间 %s ~ %s", + "[%s] %-40s → %-35s 写入 %d 点，预测区间 %s ~ %s，标签=%s", now_str, query, pred_metric, len(y_pred), future_start, future_end, + labels_to_str(write_labels), ) def main(): logger.info( - "预测服务启动 VM=%s 历史窗口=%dmin 预测窗口=%ds 轮询间隔=%ds 清理旧预测=%s", + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds 清理旧预测=%s", VM_URL, HISTORY_MINUTES, HORIZON_SECONDS, + WRITE_HORIZON_SECONDS, POLL_INTERVAL, CLEAR_OLD_PREDICTIONS, ) From 72d5c092018c5caf59f1a9f6ae556e6eff24ecca Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 20 May 2026 21:35:15 +0800 Subject: [PATCH 21/36] fix --- ai/predict_v2.py | 368 ++++++++++++++++++++++++----------------------- 1 file changed, 188 insertions(+), 180 deletions(-) diff --git a/ai/predict_v2.py b/ai/predict_v2.py index f631e12..933a34f 100755 --- a/ai/predict_v2.py +++ b/ai/predict_v2.py @@ -1,14 +1,14 @@ # -*- coding: utf-8 -*- """ -ProtoForge 预测服务 v4 +ProtoForge 预测服务 v5 修复点： -1. 不再使用 VictoriaMetrics delete_series，避免预测历史被整条删除。 -2. 不再每 30 秒写未来 120 秒，避免多轮预测窗口重叠导致 Grafana 出现竖线/毛刺。 -3. 每轮只写未来 min(HORIZON_SECONDS, POLL_INTERVAL) 秒的数据。 -4. 使用 forecast="rolling_v2" 新标签，避免和上一版 forecast="latest" 的旧预测数据混在一起。 -5. 使用真实 timestamp 做拟合，不假设采样严格等间隔。 -6. 拟合失败时不再简单写平直线，而是尽量重复最近一个周期的波形。 +1. 不再使用“单正弦拟合”作为主预测算法。 +2. 主算法改为：周期模板预测（同相位历史值加权平均）。 +3. 周期估计使用 FFT 粗估 + 自相关细化，比单纯 FFT 更稳。 +4. 若可用完整周期不足，则降级为多谐波回归（而不是单正弦）。 +5. 每轮只写入未来 min(HORIZON_SECONDS, POLL_INTERVAL) 秒，避免预测窗口重叠。 +6. 不删除旧预测历史，避免历史预测消失。 """ import logging @@ -20,17 +20,13 @@ import numpy as np import requests -from scipy.optimize import curve_fit - logging.basicConfig( level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s", ) - logger = logging.getLogger(__name__) - # ── 配置 ────────────────────────────────────────────────────────────────────── VM_URL = "http://localhost:8428" @@ -45,45 +41,36 @@ ] HISTORY_MINUTES = 30 - -# 理论预测窗口 HORIZON_SECONDS = 120 - -# 轮询间隔 POLL_INTERVAL = 30 - -# 实际写入窗口。 -# 关键点：实际写入窗口不要大于轮询间隔，否则不同批次预测会重叠。 WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) - MIN_POINTS = 120 QUERY_STEP = "1s" -# 不要再清理旧预测，否则历史预测会被整条删除。 -CLEAR_OLD_PREDICTIONS = False +# 至少要有多少个完整周期，才使用“周期模板预测” +MIN_FULL_CYCLES_FOR_TEMPLATE = 3 +MAX_CYCLES_FOR_TEMPLATE = 6 + +# 周期范围 +MIN_PERIOD_SECONDS = 5 +MAX_PERIOD_SECONDS = 3600 + +# 多谐波回归最高阶数（降级模式） +MAX_HARMONICS = 4 -# 使用新标签，避免和上一版 forecast="latest" 数据混在一起。 EXTRA_PREDICT_LABELS = { - "forecast": "rolling_v2", + "forecast": "seasonal_v1", "source": "protoforge", } -MIN_PERIOD_SECONDS = 5.0 -MAX_PERIOD_SECONDS = 3600.0 - -# 进程内记录每条预测序列上次写到哪里，避免本进程运行期间重复写同一时间段 +# 进程内记录每条预测序列上次写到哪里，避免本进程运行时重复写 LAST_WRITTEN_UNTIL: Dict[str, int] = {} # ───────────────────────────────────────────────────────────────────────────── def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: - """ - 从 VictoriaMetrics 拉取历史时序数据。 - 返回： - timestamps: Unix 秒级时间戳 - values: float 数值 - """ + """从 VictoriaMetrics 拉取历史时序数据。""" now = datetime.now() start = now - timedelta(minutes=minutes) @@ -118,20 +105,16 @@ def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[floa ts = [] ys = [] - for item in values: if len(item) < 2: continue - try: t = float(item[0]) y = float(item[1]) except Exception: continue - if not math.isfinite(t) or not math.isfinite(y): continue - ts.append(t) ys.append(y) @@ -141,33 +124,29 @@ def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[floa def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: """ 清洗历史数据： - 1. 时间戳转为整秒 + 1. 时间戳整秒化 2. 排序 - 3. 同一秒多个值时保留最后一个 - 4. 插值补齐缺失秒 + 3. 同一秒多个点保留最后一个 + 4. 按 1 秒插值补齐 """ if not ts or not ys or len(ts) != len(ys): return np.array([]), np.array([]) data = {} - for t, y in zip(ts, ys): try: sec = int(round(float(t))) val = float(y) except Exception: continue - if not math.isfinite(sec) or not math.isfinite(val): continue - data[sec] = val if not data: return np.array([]), np.array([]) sorted_items = sorted(data.items(), key=lambda x: x[0]) - ts_clean = np.array([x[0] for x in sorted_items], dtype=float) ys_clean = np.array([x[1] for x in sorted_items], dtype=float) @@ -186,22 +165,13 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. return ts_grid, ys_grid -def _sine_model(t: np.ndarray, A: float, T: float, phi: float, C: float) -> np.ndarray: - return A * np.sin(2.0 * np.pi / T * t + phi) + C - - def estimate_period_by_fft(ys_arr: np.ndarray) -> float: - """ - 使用 FFT 估算主周期。 - ys_arr 默认已经是 1 秒间隔。 - """ + """FFT 粗估周期。""" n = len(ys_arr) - - if n < 4: + if n < 8: return 60.0 centered = ys_arr - np.mean(ys_arr) - if np.allclose(centered, 0): return 60.0 @@ -212,55 +182,139 @@ def estimate_period_by_fft(ys_arr: np.ndarray) -> float: return 60.0 power = np.abs(fft_vals[1:]) - if len(power) == 0 or np.max(power) <= 0: return 60.0 dominant_idx = int(np.argmax(power)) + 1 dominant_freq = float(freqs[dominant_idx]) - if dominant_freq <= 0: return 60.0 period = 1.0 / dominant_freq - period = float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + return float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: + """ + 用自相关在 init_period 附近细化周期估计。 + """ + n = len(ys_arr) + if n < 20: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + centered = ys_arr - np.mean(ys_arr) + if np.allclose(centered, 0): + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) - return period + corr = np.correlate(centered, centered, mode="full")[n - 1:] + p0 = int(round(init_period)) + left = max(MIN_PERIOD_SECONDS, int(max(2, p0 * 0.7))) + right = min(n // 2, int(max(left + 1, p0 * 1.3))) -def repeat_last_period( - ts_grid: np.ndarray, - ys_grid: np.ndarray, - ts_future_arr: np.ndarray, - period_seconds: float, -) -> np.ndarray: + if right <= left: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + search = corr[left:right + 1] + if len(search) == 0: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + best_lag = left + int(np.argmax(search)) + return float(np.clip(best_lag, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def estimate_period(ys_arr: np.ndarray) -> float: + """FFT + 自相关 的组合周期估计。""" + p_fft = estimate_period_by_fft(ys_arr) + p_refined = refine_period_by_autocorr(ys_arr, p_fft) + return p_refined + + +def seasonal_template_predict( + ys_arr: np.ndarray, + horizon: int, + period: int, + gap: int = 0, + max_cycles: int = MAX_CYCLES_FOR_TEMPLATE, +) -> List[float]: """ - 拟合失败时的降级策略： - 不直接写平直线，而是把未来时间映射回最近一个周期的历史波形。 + 同相位历史值加权平均预测。 + 对未来第 k 个点，取过去多个周期同相位点做加权平均： + y[n-1+gap+k] ≈ avg(y[n-1+gap+k-p], y[n-1+gap+k-2p], ...) """ - if len(ts_grid) < 2: - return np.full_like(ts_future_arr, float(ys_grid[-1]), dtype=float) + n = len(ys_arr) + preds = [] - period = max(int(round(period_seconds)), 1) + for k in range(1, horizon + 1): + target_idx = (n - 1) + gap + k - y_pred = [] + values = [] + weights = [] - hist_start = float(ts_grid[0]) - hist_end = float(ts_grid[-1]) + # m=1 表示最近一个周期；m 越大越久远 + for m in range(1, max_cycles + 1): + hist_idx = target_idx - m * period + if 0 <= hist_idx < n: + # 越近权重越大 + w = 1.0 / m + values.append(float(ys_arr[hist_idx])) + weights.append(w) - for future_ts in ts_future_arr: - mapped_ts = float(future_ts) + if not values: + # 万一拿不到，退化为最后一个值 + preds.append(float(ys_arr[-1])) + else: + preds.append(float(np.average(values, weights=weights))) - while mapped_ts > hist_end: - mapped_ts -= period + return preds - while mapped_ts < hist_start: - mapped_ts += period - val = float(np.interp(mapped_ts, ts_grid, ys_grid)) - y_pred.append(val) +def harmonic_regression_predict( + ys_arr: np.ndarray, + horizon: int, + period: int, + gap: int = 0, + max_harmonics: int = MAX_HARMONICS, +) -> List[float]: + """ + 多谐波回归（降级模式）： + y = c + Σ [a_k sin(2πkt/P) + b_k cos(2πkt/P)] + 相比单正弦，更能表达非标准正弦波形。 + """ + n = len(ys_arr) + if n < 10 or period <= 1: + return [float(ys_arr[-1])] * horizon + + # 周期太短时，谐波数不能太大 + K = min(max_harmonics, max(1, period // 4)) + + t = np.arange(n, dtype=float) + cols = [np.ones(n, dtype=float)] + + for k in range(1, K + 1): + angle = 2.0 * np.pi * k * t / period + cols.append(np.sin(angle)) + cols.append(np.cos(angle)) + + X = np.column_stack(cols) + + try: + coef, _, _, _ = np.linalg.lstsq(X, ys_arr, rcond=None) + except Exception: + return [float(ys_arr[-1])] * horizon + + t_future = np.arange(n + gap, n + gap + horizon, dtype=float) + cols_future = [np.ones(horizon, dtype=float)] - return np.array(y_pred, dtype=float) + for k in range(1, K + 1): + angle = 2.0 * np.pi * k * t_future / period + cols_future.append(np.sin(angle)) + cols_future.append(np.cos(angle)) + + X_future = np.column_stack(cols_future) + y_pred = X_future @ coef + + return y_pred.astype(float).tolist() def predict_next( @@ -270,96 +324,74 @@ def predict_next( base_ts: int, ) -> Tuple[List[float], List[float]]: """ - 用 FFT 检测主频，拟合正弦波，外推未来 horizon 秒。 - - base_ts: - 从 base_ts + 1 开始写预测。 + 主预测函数： + 1. 周期估计 + 2. 优先使用周期模板预测 + 3. 周期不够时降级为多谐波回归 """ ts_grid, ys_grid = normalize_history(ts, ys) - if len(ys_grid) < MIN_POINTS: return [], [] y_min = float(np.min(ys_grid)) y_max = float(np.max(ys_grid)) - y_mean = float(np.mean(ys_grid)) y_range = y_max - y_min - base_ts = max(int(base_ts), int(ts_grid[-1])) - - ts_future_arr = np.arange( - base_ts + 1, - base_ts + 1 + horizon, - 1, - dtype=float, - ) - if y_range <= 1e-9: - y_pred_arr = np.full_like(ts_future_arr, float(ys_grid[-1]), dtype=float) - return ts_future_arr.tolist(), y_pred_arr.tolist() + base_ts = max(int(base_ts), int(ts_grid[-1])) + ts_future = [base_ts + i + 1 for i in range(horizon)] + y_pred = [float(ys_grid[-1])] * horizon + return ts_future, y_pred - period = estimate_period_by_fft(ys_grid) + period_est = estimate_period(ys_grid) + period = int(round(period_est)) + period = max(MIN_PERIOD_SECONDS, min(MAX_PERIOD_SECONDS, period)) - t_fit = ts_grid - ts_grid[0] - t_future = ts_future_arr - ts_grid[0] + last_real_ts = int(ts_grid[-1]) + base_ts = max(int(base_ts), last_real_ts) - amplitude = y_range / 2.0 - offset = y_mean + # 如果当前时间已经超过最后一个真实点，gap 表示中间“空过去”的秒数 + gap = max(0, base_ts - last_real_ts) - try: - popt, _ = curve_fit( - _sine_model, - t_fit, - ys_grid, - p0=[amplitude, period, 0.0, offset], - bounds=( - [0.0, MIN_PERIOD_SECONDS, -2.0 * np.pi, y_min - y_range], - [np.inf, MAX_PERIOD_SECONDS, 2.0 * np.pi, y_max + y_range], - ), - maxfev=12000, - ) + ts_future = [base_ts + i + 1 for i in range(horizon)] - y_pred_arr = _sine_model(t_future, *popt) + full_cycles = len(ys_grid) // period if period > 0 else 0 - margin = y_range * 0.2 - lower = y_min - margin - upper = y_max + margin - y_pred_arr = np.clip(y_pred_arr, lower, upper) - - if not np.all(np.isfinite(y_pred_arr)): - raise ValueError("预测结果包含 NaN/Inf") - - logger.debug( - "正弦拟合成功 period=%.2fs amplitude=%.4f offset=%.4f", - popt[1], - popt[0], - popt[3], + if full_cycles >= MIN_FULL_CYCLES_FOR_TEMPLATE: + y_pred = seasonal_template_predict( + ys_arr=ys_grid, + horizon=horizon, + period=period, + gap=gap, + max_cycles=min(MAX_CYCLES_FOR_TEMPLATE, full_cycles), ) - - return ts_future_arr.tolist(), y_pred_arr.astype(float).tolist() - - except Exception as e: - logger.warning("正弦拟合失败，降级为最近周期波形复制: %s", e) - - y_pred_arr = repeat_last_period( - ts_grid=ts_grid, - ys_grid=ys_grid, - ts_future_arr=ts_future_arr, - period_seconds=period, + model_name = "seasonal_template" + else: + y_pred = harmonic_regression_predict( + ys_arr=ys_grid, + horizon=horizon, + period=period, + gap=gap, + max_harmonics=MAX_HARMONICS, ) + model_name = "harmonic_regression" + + # 合理裁剪，避免偶然外推过大 + margin = y_range * 0.15 + lower = y_min - margin + upper = y_max + margin + y_pred = np.clip(np.array(y_pred, dtype=float), lower, upper).astype(float).tolist() - margin = y_range * 0.2 - lower = y_min - margin - upper = y_max + margin - y_pred_arr = np.clip(y_pred_arr, lower, upper) + logger.debug( + "predict_next model=%s period=%ss full_cycles=%s gap=%s", + model_name, period, full_cycles, gap + ) - return ts_future_arr.tolist(), y_pred_arr.astype(float).tolist() + return ts_future, y_pred def prom_escape_label_value(value: str) -> str: - """ - Prometheus exposition label value 转义。 - """ + """Prometheus label value 转义。""" return ( str(value) .replace("\\", "\\\\") @@ -371,13 +403,10 @@ def prom_escape_label_value(value: str) -> str: def labels_to_str(labels: Dict[str, str]) -> str: if not labels: return "" - parts = [] - for k in sorted(labels.keys()): v = prom_escape_label_value(labels[k]) parts.append(f'{k}="{v}"') - return "{" + ",".join(parts) + "}" @@ -387,16 +416,12 @@ def write_predictions( metric_name: str, labels: Dict[str, str], ) -> bool: - """ - 将预测值以 Prometheus exposition 格式写入 VictoriaMetrics。 - 时间戳为毫秒级 Unix timestamp。 - """ + """将预测值以 Prometheus exposition 格式写入 VictoriaMetrics。""" if not ts_future or not y_pred or len(ts_future) != len(y_pred): logger.warning("预测数据为空或长度不一致 metric=%s", metric_name) return False label_str = labels_to_str(labels) - lines = [] for t, y in zip(ts_future, y_pred): @@ -422,14 +447,11 @@ def write_predictions( resp = requests.post( f"{VM_URL}/api/v1/import/prometheus", data=payload.encode("utf-8"), - headers={ - "Content-Type": "text/plain; version=0.0.4; charset=utf-8", - }, + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, timeout=10, ) resp.raise_for_status() return True - except requests.RequestException as e: logger.error("写入预测数据失败 metric=%s: %s", metric_name, e) return False @@ -441,12 +463,7 @@ def write_predictions( def _parse_labels(query: str) -> Dict[str, str]: - """ - 从查询表达式中解析标签。 - - 示例： - feed_rate{device_id="fanuc-cnc"} -> {"device_id": "fanuc-cnc"} - """ + """从查询表达式中解析标签。""" labels = {} if "{" not in query or "}" not in query: @@ -468,19 +485,13 @@ def _parse_labels(query: str) -> Dict[str, str]: def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: result = {} - for d in dicts: - if not d: - continue - result.update(d) - + if d: + result.update(d) return result def series_key(metric_name: str, labels: Dict[str, str]) -> str: - """ - 构造进程内唯一 key，用于记录上次写到哪个时间点。 - """ return metric_name + labels_to_str(labels) @@ -489,7 +500,6 @@ def run_once(): for query, pred_metric in PREDICT_TARGETS: ts, ys = fetch_history(query) - if len(ys) < MIN_POINTS: logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) continue @@ -502,7 +512,7 @@ def run_once(): now_sec = int(time.time()) last_until = LAST_WRITTEN_UNTIL.get(key, 0) - # 防止同一进程内重复写入已经预测过的时间段 + # 避免同一进程内写重叠时间段 base_ts = max(now_sec, last_until) ts_future, y_pred = predict_next( @@ -522,7 +532,6 @@ def run_once(): metric_name=pred_metric, labels=write_labels, ) - if not ok: continue @@ -545,13 +554,12 @@ def run_once(): def main(): logger.info( - "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds 清理旧预测=%s", + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds", VM_URL, HISTORY_MINUTES, HORIZON_SECONDS, WRITE_HORIZON_SECONDS, POLL_INTERVAL, - CLEAR_OLD_PREDICTIONS, ) while True: From c26b9991d4cfac374d56829b47023b043d270aba Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 09:03:00 +0800 Subject: [PATCH 22/36] feat(ai): support single scene predict --- ai/predict_v3_single_scene.py | 1058 +++++++++++++++++++++++++++++++++ 1 file changed, 1058 insertions(+) create mode 100644 ai/predict_v3_single_scene.py diff --git a/ai/predict_v3_single_scene.py b/ai/predict_v3_single_scene.py new file mode 100644 index 0000000..23af8c5 --- /dev/null +++ b/ai/predict_v3_single_scene.py @@ -0,0 +1,1058 @@ +# -*- coding: utf-8 -*- +""" +ProtoForge 预测服务 v6 + +核心能力： +1. 周期模板预测：适合 CNC 这类强周期、非标准正弦波形。 +2. 健康基线冻结：检测到异常后，不再用故障数据更新预测模板。 +3. 恢复冷却机制：故障恢复后，需要连续稳定多个周期，才恢复学习。 +4. 预测上下界：写入 predicted_upper / predicted_lower，方便 Grafana 展示预测带。 +5. 异常标记：写入 xxx_anomaly，1 表示异常，0 表示正常。 +6. 不删除历史预测，不使用 delete_series。 +""" + +""" +场景：不考虑物料、不考虑跨程序场景算法预测 +""" + +import json +import logging +import math +import os +import re +import time +from dataclasses import asdict, dataclass +from datetime import datetime, timedelta +from typing import Dict, List, Optional, Tuple + +import numpy as np +import requests + + +logging.basicConfig( + level=logging.INFO, + format="%(asctime)s [%(levelname)s] %(message)s", +) + +logger = logging.getLogger(__name__) + + +# ── 基础配置 ────────────────────────────────────────────────────────────────── + +VM_URL = "http://localhost:8428" + +STATE_FILE = "/tmp/protoforge_predictor_state.json" + +HISTORY_MINUTES = 30 +HORIZON_SECONDS = 120 +POLL_INTERVAL = 30 + +# 实际每轮写入未来多少秒。 +# 不要大于 POLL_INTERVAL，否则多轮预测会重叠。 +WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) + +QUERY_STEP = "1s" +MIN_POINTS = 120 + +MIN_PERIOD_SECONDS = 5 +MAX_PERIOD_SECONDS = 3600 + +# 至少多少个完整周期才允许构建健康模板 +MIN_FULL_CYCLES_FOR_TEMPLATE = 3 + +# 构建模板最多使用最近多少个周期 +MAX_CYCLES_FOR_TEMPLATE = 6 + +# 检测异常使用最近多少秒实际数据 +DETECT_WINDOW_SECONDS = 15 + +# 恢复后，至少连续正常多少秒才考虑恢复学习 +RECOVERY_MIN_SECONDS = 60 + +# 健康状态下模板更新速度，越小越保守 +HEALTHY_EMA_ALPHA = 0.15 + +# 故障恢复后第一次重新学习时的更新速度 +RECOVERY_EMA_ALPHA = 0.35 + +# 最近窗口里有多少比例的点超过阈值，才认为异常 +OUTSIDE_RATIO_THRESHOLD = 0.60 + +# 最近窗口里有多少比例的点回到阈值内，才认为恢复正常 +RECOVERY_INSIDE_RATIO_THRESHOLD = 0.80 + + +# ── 指标配置 ────────────────────────────────────────────────────────────────── +# abs_threshold / rel_threshold 需要按指标单位调。 +# feed_rate 单位 mm/min，这里先给 400 和 25%。 + +PREDICT_TARGETS = [ + { + "query": 'feed_rate{device_id="fanuc-cnc"}', + "pred_metric": "feed_rate_predicted", + "anomaly_metric": "feed_rate_anomaly", + "abs_threshold": 400.0, + "rel_threshold": 0.25, + }, + { + "query": 'spindle_speed{device_id="fanuc-cnc"}', + "pred_metric": "spindle_speed_predicted", + "anomaly_metric": "spindle_speed_anomaly", + "abs_threshold": 500.0, + "rel_threshold": 0.25, + }, + { + "query": 'spindle_current{device_id="fanuc-cnc"}', + "pred_metric": "spindle_current_predicted", + "anomaly_metric": "spindle_current_anomaly", + "abs_threshold": 5.0, + "rel_threshold": 0.25, + }, + { + "query": 'vibration_x{device_id="fanuc-cnc"}', + "pred_metric": "vibration_x_predicted", + "anomaly_metric": "vibration_x_anomaly", + "abs_threshold": 1.0, + "rel_threshold": 0.30, + }, + { + "query": 'vibration_y{device_id="fanuc-cnc"}', + "pred_metric": "vibration_y_predicted", + "anomaly_metric": "vibration_y_anomaly", + "abs_threshold": 1.0, + "rel_threshold": 0.30, + }, + { + "query": 'vibration_z{device_id="fanuc-cnc"}', + "pred_metric": "vibration_z_predicted", + "anomaly_metric": "vibration_z_anomaly", + "abs_threshold": 1.0, + "rel_threshold": 0.30, + }, +] + +EXTRA_PREDICT_LABELS = { + "forecast": "health_gated_v1", + "source": "protoforge", +} + +BASELINE_STATUS_HEALTHY = "healthy" +BASELINE_STATUS_ANOMALY = "anomaly" +BASELINE_STATUS_RECOVERING = "recovering" +BASELINE_STATUS_LEARNING = "learning" + + +# ── 状态结构 ────────────────────────────────────────────────────────────────── + +@dataclass +class BaselineState: + period: int + template: List[float] + status: str + clean_seconds: int + last_update_ts: int + last_seen_ts: int + y_min: float + y_max: float + + +BASELINE_STATES: Dict[str, BaselineState] = {} +LAST_WRITTEN_UNTIL: Dict[str, int] = {} + + +# ── VM 读取 ─────────────────────────────────────────────────────────────────── + +def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: + now = datetime.now() + start = now - timedelta(minutes=minutes) + + try: + resp = requests.get( + f"{VM_URL}/api/v1/query_range", + params={ + "query": query, + "start": start.timestamp(), + "end": now.timestamp(), + "step": QUERY_STEP, + }, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("拉取数据失败 query=%s: %s", query, e) + return [], [] + + try: + result = resp.json().get("data", {}).get("result", []) + except Exception as e: + logger.error("解析 VM 返回失败 query=%s: %s", query, e) + return [], [] + + if not result: + return [], [] + + values = result[0].get("values", []) + if not values: + return [], [] + + ts = [] + ys = [] + + for item in values: + if len(item) < 2: + continue + + try: + t = float(item[0]) + y = float(item[1]) + except Exception: + continue + + if not math.isfinite(t) or not math.isfinite(y): + continue + + ts.append(t) + ys.append(y) + + return ts, ys + + +def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: + if not ts or not ys or len(ts) != len(ys): + return np.array([]), np.array([]) + + data = {} + + for t, y in zip(ts, ys): + try: + sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(sec) or not math.isfinite(val): + continue + + data[sec] = val + + if not data: + return np.array([]), np.array([]) + + sorted_items = sorted(data.items(), key=lambda x: x[0]) + + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) + ys_clean = np.array([x[1] for x in sorted_items], dtype=float) + + if len(ts_clean) < 2: + return ts_clean, ys_clean + + start_sec = int(ts_clean[0]) + end_sec = int(ts_clean[-1]) + + if end_sec <= start_sec: + return ts_clean, ys_clean + + ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) + ys_grid = np.interp(ts_grid, ts_clean, ys_clean) + + return ts_grid, ys_grid + + +# ── 周期估计 ────────────────────────────────────────────────────────────────── + +def estimate_period_by_fft(ys_arr: np.ndarray) -> float: + n = len(ys_arr) + + if n < 8: + return 60.0 + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return 60.0 + + fft_vals = np.fft.rfft(centered) + freqs = np.fft.rfftfreq(n, d=1.0) + + if len(freqs) <= 1: + return 60.0 + + power = np.abs(fft_vals[1:]) + + if len(power) == 0 or np.max(power) <= 0: + return 60.0 + + dominant_idx = int(np.argmax(power)) + 1 + dominant_freq = float(freqs[dominant_idx]) + + if dominant_freq <= 0: + return 60.0 + + period = 1.0 / dominant_freq + + return float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: + n = len(ys_arr) + + if n < 20: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + corr = np.correlate(centered, centered, mode="full")[n - 1:] + + p0 = int(round(init_period)) + left = max(MIN_PERIOD_SECONDS, int(max(2, p0 * 0.7))) + right = min(n // 2, int(max(left + 1, p0 * 1.3))) + + if right <= left: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + search = corr[left:right + 1] + + if len(search) == 0: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + best_lag = left + int(np.argmax(search)) + + return float(np.clip(best_lag, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def estimate_period(ys_arr: np.ndarray) -> int: + p_fft = estimate_period_by_fft(ys_arr) + p_refined = refine_period_by_autocorr(ys_arr, p_fft) + + period = int(round(p_refined)) + period = max(MIN_PERIOD_SECONDS, min(MAX_PERIOD_SECONDS, period)) + + return int(period) + + +# ── 模板构建与预测 ───────────────────────────────────────────────────────────── + +def fill_template_nan(template: np.ndarray) -> np.ndarray: + period = len(template) + + if period == 0: + return template + + idx = np.arange(period) + valid = np.isfinite(template) + + if not np.any(valid): + return np.zeros(period, dtype=float) + + if np.all(valid): + return template + + x_valid = idx[valid] + y_valid = template[valid] + + # 环形插值，处理 phase 0 附近缺口 + x_ext = np.concatenate([x_valid - period, x_valid, x_valid + period]) + y_ext = np.concatenate([y_valid, y_valid, y_valid]) + + filled = np.interp(idx, x_ext, y_ext) + + return filled.astype(float) + + +def build_phase_template( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + period: int, + max_cycles: int = MAX_CYCLES_FOR_TEMPLATE, + tail_seconds: Optional[int] = None, +) -> Optional[np.ndarray]: + if period <= 1 or len(ys_grid) < period * MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + max_seconds = period * max_cycles + + if tail_seconds is not None: + max_seconds = min(max_seconds, int(tail_seconds)) + + max_seconds = max(period * MIN_FULL_CYCLES_FOR_TEMPLATE, max_seconds) + + if len(ys_grid) < max_seconds: + start_idx = 0 + else: + start_idx = len(ys_grid) - max_seconds + + ts_tail = ts_grid[start_idx:] + ys_tail = ys_grid[start_idx:] + + if len(ys_tail) < period * MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + sums = np.zeros(period, dtype=float) + weights = np.zeros(period, dtype=float) + + total = len(ys_tail) + + for i, (t, y) in enumerate(zip(ts_tail, ys_tail)): + phase = int(t) % period + + # 越近的数据权重越高 + recency = (i + 1) / total + weight = 0.3 + 0.7 * recency + + sums[phase] += float(y) * weight + weights[phase] += weight + + template = np.full(period, np.nan, dtype=float) + + valid = weights > 0 + template[valid] = sums[valid] / weights[valid] + + template = fill_template_nan(template) + + return template + + +def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: + old_period = len(old_template) + + if old_period == new_period: + return old_template.astype(float) + + if old_period <= 1 or new_period <= 1: + return np.full(new_period, float(np.mean(old_template)), dtype=float) + + old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) + new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) + + old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) + old_y_ext = np.concatenate([old_template, old_template, old_template]) + + return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + + +def merge_template( + old_template: np.ndarray, + new_template: np.ndarray, + alpha: float, +) -> np.ndarray: + alpha = float(np.clip(alpha, 0.0, 1.0)) + + if len(old_template) != len(new_template): + old_template = resample_template(old_template, len(new_template)) + + return ((1.0 - alpha) * old_template + alpha * new_template).astype(float) + + +def predict_by_state(state: BaselineState, ts_list: List[int]) -> np.ndarray: + template = np.array(state.template, dtype=float) + period = int(state.period) + + if period <= 1 or len(template) != period: + return np.zeros(len(ts_list), dtype=float) + + values = [] + + for ts in ts_list: + phase = int(ts) % period + values.append(float(template[phase])) + + return np.array(values, dtype=float) + + +def calc_threshold(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> np.ndarray: + return np.maximum(abs_threshold, np.abs(pred) * rel_threshold) + + +def calc_bounds(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> Tuple[np.ndarray, np.ndarray]: + threshold = calc_threshold(pred, abs_threshold, rel_threshold) + lower = pred - threshold + upper = pred + threshold + return lower, upper + + +# ── 异常检测与状态更新 ──────────────────────────────────────────────────────── + +def detect_anomaly( + state: BaselineState, + ts_grid: np.ndarray, + ys_grid: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[bool, float, float, float]: + if len(ys_grid) < DETECT_WINDOW_SECONDS: + return False, 0.0, 0.0, 0.0 + + ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() + actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) + + pred = predict_by_state(state, ts_recent) + threshold = calc_threshold(pred, abs_threshold, rel_threshold) + + abs_err = np.abs(actual - pred) + outside = abs_err > threshold + + outside_ratio = float(np.mean(outside)) + mean_abs_err = float(np.mean(abs_err)) + mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1.0))) + + is_anomaly = outside_ratio >= OUTSIDE_RATIO_THRESHOLD + + return is_anomaly, outside_ratio, mean_abs_err, mean_rel_err + + +def is_recovered( + state: BaselineState, + ts_grid: np.ndarray, + ys_grid: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[bool, float]: + if len(ys_grid) < DETECT_WINDOW_SECONDS: + return False, 0.0 + + ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() + actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) + + pred = predict_by_state(state, ts_recent) + threshold = calc_threshold(pred, abs_threshold, rel_threshold) + + abs_err = np.abs(actual - pred) + inside = abs_err <= threshold + + inside_ratio = float(np.mean(inside)) + + return inside_ratio >= RECOVERY_INSIDE_RATIO_THRESHOLD, inside_ratio + + +def create_initial_state( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + now_sec: int, +) -> Optional[BaselineState]: + if len(ys_grid) < MIN_POINTS: + return None + + period = estimate_period(ys_grid) + + template = build_phase_template( + ts_grid=ts_grid, + ys_grid=ys_grid, + period=period, + max_cycles=MAX_CYCLES_FOR_TEMPLATE, + tail_seconds=period * MAX_CYCLES_FOR_TEMPLATE, + ) + + if template is None: + return None + + return BaselineState( + period=int(period), + template=template.astype(float).tolist(), + status=BASELINE_STATUS_HEALTHY, + clean_seconds=int(period * MAX_CYCLES_FOR_TEMPLATE), + last_update_ts=now_sec, + last_seen_ts=now_sec, + y_min=float(np.min(ys_grid)), + y_max=float(np.max(ys_grid)), + ) + + +def maybe_update_state( + key: str, + ts_grid: np.ndarray, + ys_grid: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[Optional[BaselineState], bool, float, float, float]: + now_sec = int(time.time()) + + state = BASELINE_STATES.get(key) + + if state is None: + state = create_initial_state(ts_grid, ys_grid, now_sec) + + if state is None: + return None, False, 0.0, 0.0, 0.0 + + BASELINE_STATES[key] = state + logger.info( + "初始化健康模板 key=%s period=%ss clean_seconds=%ss", + key, + state.period, + state.clean_seconds, + ) + return state, False, 0.0, 0.0, 0.0 + + elapsed = max(1, now_sec - int(state.last_seen_ts)) + elapsed = min(elapsed, POLL_INTERVAL * 2) + state.last_seen_ts = now_sec + + is_anom, outside_ratio, mean_abs_err, mean_rel_err = detect_anomaly( + state=state, + ts_grid=ts_grid, + ys_grid=ys_grid, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + ) + + if is_anom: + state.status = BASELINE_STATUS_ANOMALY + state.clean_seconds = 0 + + logger.warning( + "检测到异常，冻结模板 key=%s outside_ratio=%.2f mean_abs_err=%.2f mean_rel_err=%.2f", + key, + outside_ratio, + mean_abs_err, + mean_rel_err, + ) + + BASELINE_STATES[key] = state + return state, True, outside_ratio, mean_abs_err, mean_rel_err + + recovered, inside_ratio = is_recovered( + state=state, + ts_grid=ts_grid, + ys_grid=ys_grid, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + ) + + if state.status == BASELINE_STATUS_ANOMALY: + if recovered: + state.status = BASELINE_STATUS_RECOVERING + state.clean_seconds = elapsed + logger.info( + "异常开始恢复 key=%s inside_ratio=%.2f clean_seconds=%ss", + key, + inside_ratio, + state.clean_seconds, + ) + else: + state.clean_seconds = 0 + BASELINE_STATES[key] = state + return state, True, outside_ratio, mean_abs_err, mean_rel_err + + elif state.status == BASELINE_STATUS_RECOVERING: + if recovered: + state.clean_seconds += elapsed + else: + state.status = BASELINE_STATUS_ANOMALY + state.clean_seconds = 0 + BASELINE_STATES[key] = state + return state, True, outside_ratio, mean_abs_err, mean_rel_err + + else: + state.status = BASELINE_STATUS_HEALTHY + state.clean_seconds += elapsed + + # 故障恢复后，不要立刻学习。 + # 必须至少连续正常：max(RECOVERY_MIN_SECONDS, 3 个周期) + min_clean_for_update = max( + RECOVERY_MIN_SECONDS, + int(state.period) * MIN_FULL_CYCLES_FOR_TEMPLATE, + ) + + if state.clean_seconds < min_clean_for_update: + BASELINE_STATES[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + # 只使用最近 clean_seconds 这段连续正常数据来更新模板，避免历史故障污染。 + new_period = estimate_period(ys_grid) + tail_seconds = min( + int(state.clean_seconds), + int(new_period) * MAX_CYCLES_FOR_TEMPLATE, + ) + + new_template = build_phase_template( + ts_grid=ts_grid, + ys_grid=ys_grid, + period=new_period, + max_cycles=MAX_CYCLES_FOR_TEMPLATE, + tail_seconds=tail_seconds, + ) + + if new_template is None: + BASELINE_STATES[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + old_template = np.array(state.template, dtype=float) + + if state.status == BASELINE_STATUS_RECOVERING: + alpha = RECOVERY_EMA_ALPHA + state.status = BASELINE_STATUS_HEALTHY + else: + alpha = HEALTHY_EMA_ALPHA + + merged = merge_template( + old_template=old_template, + new_template=new_template, + alpha=alpha, + ) + + state.period = int(new_period) + state.template = merged.astype(float).tolist() + state.last_update_ts = now_sec + state.y_min = float(np.min(ys_grid[-tail_seconds:])) + state.y_max = float(np.max(ys_grid[-tail_seconds:])) + + BASELINE_STATES[key] = state + + logger.info( + "更新健康模板 key=%s period=%ss status=%s clean_seconds=%ss alpha=%.2f", + key, + state.period, + state.status, + state.clean_seconds, + alpha, + ) + + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + +# ── Prometheus 格式写入 ─────────────────────────────────────────────────────── + +def prom_escape_label_value(value: str) -> str: + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) + + +def labels_to_str(labels: Dict[str, str]) -> str: + if not labels: + return "" + + parts = [] + + for k in sorted(labels.keys()): + v = prom_escape_label_value(labels[k]) + parts.append(f'{k}="{v}"') + + return "{" + ",".join(parts) + "}" + + +def write_series( + metric_name: str, + labels: Dict[str, str], + ts_list: List[int], + values: List[float], +) -> bool: + if not ts_list or not values or len(ts_list) != len(values): + return False + + label_str = labels_to_str(labels) + lines = [] + + for t, y in zip(ts_list, values): + try: + ts_sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(ts_sec) or not math.isfinite(val): + continue + + ts_ms = ts_sec * 1000 + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_ms}") + + if not lines: + return False + + payload = "\n".join(lines) + "\n" + + try: + resp = requests.post( + f"{VM_URL}/api/v1/import/prometheus", + data=payload.encode("utf-8"), + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + timeout=10, + ) + resp.raise_for_status() + return True + except requests.RequestException as e: + logger.error("写入数据失败 metric=%s: %s", metric_name, e) + return False + + +def write_prediction_bundle( + pred_metric: str, + anomaly_metric: str, + labels: Dict[str, str], + ts_future: List[int], + pred_values: np.ndarray, + lower_values: np.ndarray, + upper_values: np.ndarray, + is_anomaly: bool, + outside_ratio: float, + mean_abs_err: float, + mean_rel_err: float, +) -> bool: + ok1 = write_series( + metric_name=pred_metric, + labels=labels, + ts_list=ts_future, + values=pred_values.astype(float).tolist(), + ) + + ok2 = write_series( + metric_name=f"{pred_metric}_lower", + labels=labels, + ts_list=ts_future, + values=lower_values.astype(float).tolist(), + ) + + ok3 = write_series( + metric_name=f"{pred_metric}_upper", + labels=labels, + ts_list=ts_future, + values=upper_values.astype(float).tolist(), + ) + + now_sec = int(time.time()) + + anomaly_labels = dict(labels) + anomaly_labels["type"] = "prediction_deviation" + + ok4 = write_series( + metric_name=anomaly_metric, + labels=anomaly_labels, + ts_list=[now_sec], + values=[1.0 if is_anomaly else 0.0], + ) + + ok5 = write_series( + metric_name=f"{anomaly_metric}_outside_ratio", + labels=anomaly_labels, + ts_list=[now_sec], + values=[outside_ratio], + ) + + ok6 = write_series( + metric_name=f"{anomaly_metric}_mean_abs_error", + labels=anomaly_labels, + ts_list=[now_sec], + values=[mean_abs_err], + ) + + ok7 = write_series( + metric_name=f"{anomaly_metric}_mean_rel_error", + labels=anomaly_labels, + ts_list=[now_sec], + values=[mean_rel_err], + ) + + return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 + + +# ── 标签解析 ────────────────────────────────────────────────────────────────── + +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) + + +def _parse_labels(query: str) -> Dict[str, str]: + labels = {} + + if "{" not in query or "}" not in query: + return labels + + try: + label_part = query[query.index("{") + 1: query.rindex("}")] + except Exception: + return labels + + for match in _LABEL_PATTERN.finditer(label_part): + key = match.group(1) + value = match.group(2) + value = value.replace('\\"', '"').replace("\\n", "\n").replace("\\\\", "\\") + labels[key] = value + + return labels + + +def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: + result = {} + + for d in dicts: + if d: + result.update(d) + + return result + + +def series_key(metric_name: str, labels: Dict[str, str]) -> str: + return metric_name + labels_to_str(labels) + + +# ── 状态持久化 ──────────────────────────────────────────────────────────────── + +def load_state(): + global BASELINE_STATES + + if not os.path.exists(STATE_FILE): + return + + try: + with open(STATE_FILE, "r", encoding="utf-8") as f: + raw = json.load(f) + + states = {} + + for key, value in raw.get("baseline_states", {}).items(): + states[key] = BaselineState(**value) + + BASELINE_STATES = states + + logger.info("已加载预测状态文件 %s，状态数量=%d", STATE_FILE, len(BASELINE_STATES)) + + except Exception as e: + logger.warning("加载预测状态文件失败，将重新学习: %s", e) + + +def save_state(): + try: + raw = { + "baseline_states": { + key: asdict(value) + for key, value in BASELINE_STATES.items() + } + } + + tmp_file = STATE_FILE + ".tmp" + + with open(tmp_file, "w", encoding="utf-8") as f: + json.dump(raw, f, ensure_ascii=False, indent=2) + + os.replace(tmp_file, STATE_FILE) + + except Exception as e: + logger.warning("保存预测状态文件失败: %s", e) + + +# ── 主逻辑 ──────────────────────────────────────────────────────────────────── + +def run_once(): + now_str = datetime.now().strftime("%H:%M:%S") + + for target in PREDICT_TARGETS: + query = target["query"] + pred_metric = target["pred_metric"] + anomaly_metric = target["anomaly_metric"] + abs_threshold = float(target["abs_threshold"]) + rel_threshold = float(target["rel_threshold"]) + + ts, ys = fetch_history(query) + + if len(ys) < MIN_POINTS: + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + continue + + ts_grid, ys_grid = normalize_history(ts, ys) + + if len(ys_grid) < MIN_POINTS: + logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid)) + continue + + base_labels = _parse_labels(query) + write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + + key = series_key(pred_metric, write_labels) + + state, is_anomaly, outside_ratio, mean_abs_err, mean_rel_err = maybe_update_state( + key=key, + ts_grid=ts_grid, + ys_grid=ys_grid, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + ) + + if state is None: + logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + continue + + now_sec = int(time.time()) + last_until = LAST_WRITTEN_UNTIL.get(key, 0) + last_real_ts = int(ts_grid[-1]) + + base_ts = max(now_sec, last_until, last_real_ts) + + ts_future = [ + base_ts + i + 1 + for i in range(WRITE_HORIZON_SECONDS) + ] + + pred_values = predict_by_state(state, ts_future) + + lower_values, upper_values = calc_bounds( + pred=pred_values, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + ) + + ok = write_prediction_bundle( + pred_metric=pred_metric, + anomaly_metric=anomaly_metric, + labels=write_labels, + ts_future=ts_future, + pred_values=pred_values, + lower_values=lower_values, + upper_values=upper_values, + is_anomaly=is_anomaly, + outside_ratio=outside_ratio, + mean_abs_err=mean_abs_err, + mean_rel_err=mean_rel_err, + ) + + if not ok: + continue + + LAST_WRITTEN_UNTIL[key] = int(max(ts_future)) + + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") + future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + + logger.info( + "[%s] %-40s → %-35s status=%s anomaly=%s period=%ss clean=%ss 写入 %d 点，预测区间 %s ~ %s", + now_str, + query, + pred_metric, + state.status, + is_anomaly, + state.period, + state.clean_seconds, + len(ts_future), + future_start, + future_end, + ) + + save_state() + + +def main(): + load_state() + + logger.info( + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s", + VM_URL, + HISTORY_MINUTES, + HORIZON_SECONDS, + WRITE_HORIZON_SECONDS, + POLL_INTERVAL, + STATE_FILE, + ) + + while True: + run_once() + time.sleep(POLL_INTERVAL) + + +if __name__ == "__main__": + main() \ No newline at end of file From 54c4b851a004567078cfec337933aafcbd676b44 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 13:39:33 +0800 Subject: [PATCH 23/36] fix --- ai/predict_v3_single_scene.py | 701 +++++++++++++++++++++++----------- 1 file changed, 488 insertions(+), 213 deletions(-) mode change 100644 => 100755 ai/predict_v3_single_scene.py diff --git a/ai/predict_v3_single_scene.py b/ai/predict_v3_single_scene.py old mode 100644 new mode 100755 index 23af8c5..fc07f4f --- a/ai/predict_v3_single_scene.py +++ b/ai/predict_v3_single_scene.py @@ -1,18 +1,22 @@ # -*- coding: utf-8 -*- """ -ProtoForge 预测服务 v6 - -核心能力： -1. 周期模板预测：适合 CNC 这类强周期、非标准正弦波形。 -2. 健康基线冻结：检测到异常后，不再用故障数据更新预测模板。 -3. 恢复冷却机制：故障恢复后，需要连续稳定多个周期，才恢复学习。 -4. 预测上下界：写入 predicted_upper / predicted_lower，方便 Grafana 展示预测带。 -5. 异常标记：写入 xxx_anomaly，1 表示异常，0 表示正常。 -6. 不删除历史预测，不使用 delete_series。 -""" - -""" -场景：不考虑物料、不考虑跨程序场景算法预测 +ProtoForge Predictor v8 + +功能： +1. 从 VictoriaMetrics 拉取历史数据。 +2. 对 CNC 周期型指标进行相位对齐预测。 +3. 使用“谷底锚点”对齐周期，减少上升沿/下降沿相位偏差。 +4. 每轮只写入未来 min(HORIZON_SECONDS, POLL_INTERVAL) 秒，避免预测窗口重叠。 +5. 检测异常后冻结健康模板，不把故障数据学进去。 +6. 故障恢复后等待稳定一段时间，再恢复模板更新。 +7. 写入： + - xxx_predicted + - xxx_predicted_upper + - xxx_predicted_lower + - xxx_anomaly + - xxx_anomaly_outside_ratio + - xxx_anomaly_mean_abs_error + - xxx_anomaly_mean_rel_error """ import json @@ -29,6 +33,10 @@ import requests +# ============================================================================= +# 日志配置 +# ============================================================================= + logging.basicConfig( level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s", @@ -37,18 +45,19 @@ logger = logging.getLogger(__name__) -# ── 基础配置 ────────────────────────────────────────────────────────────────── +# ============================================================================= +# 基础配置 +# ============================================================================= VM_URL = "http://localhost:8428" -STATE_FILE = "/tmp/protoforge_predictor_state.json" +STATE_FILE = "/tmp/protoforge_predictor_state_v8.json" HISTORY_MINUTES = 30 HORIZON_SECONDS = 120 POLL_INTERVAL = 30 -# 实际每轮写入未来多少秒。 -# 不要大于 POLL_INTERVAL，否则多轮预测会重叠。 +# 实际写入窗口不要大于轮询间隔，否则多轮预测会重叠。 WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) QUERY_STEP = "1s" @@ -57,34 +66,25 @@ MIN_PERIOD_SECONDS = 5 MAX_PERIOD_SECONDS = 3600 -# 至少多少个完整周期才允许构建健康模板 MIN_FULL_CYCLES_FOR_TEMPLATE = 3 - -# 构建模板最多使用最近多少个周期 MAX_CYCLES_FOR_TEMPLATE = 6 -# 检测异常使用最近多少秒实际数据 DETECT_WINDOW_SECONDS = 15 - -# 恢复后，至少连续正常多少秒才考虑恢复学习 RECOVERY_MIN_SECONDS = 60 -# 健康状态下模板更新速度，越小越保守 -HEALTHY_EMA_ALPHA = 0.15 - -# 故障恢复后第一次重新学习时的更新速度 -RECOVERY_EMA_ALPHA = 0.35 +HEALTHY_EMA_ALPHA = 0.12 +RECOVERY_EMA_ALPHA = 0.30 -# 最近窗口里有多少比例的点超过阈值，才认为异常 OUTSIDE_RATIO_THRESHOLD = 0.60 - -# 最近窗口里有多少比例的点回到阈值内，才认为恢复正常 RECOVERY_INSIDE_RATIO_THRESHOLD = 0.80 +PHASE_SEARCH_RATIO = 0.15 +VALLEY_QUANTILE = 45 + -# ── 指标配置 ────────────────────────────────────────────────────────────────── -# abs_threshold / rel_threshold 需要按指标单位调。 -# feed_rate 单位 mm/min，这里先给 400 和 25%。 +# ============================================================================= +# 预测指标配置 +# ============================================================================= PREDICT_TARGETS = [ { @@ -132,21 +132,23 @@ ] EXTRA_PREDICT_LABELS = { - "forecast": "health_gated_v1", + "forecast": "phase_aligned_health_v8", "source": "protoforge", } BASELINE_STATUS_HEALTHY = "healthy" BASELINE_STATUS_ANOMALY = "anomaly" BASELINE_STATUS_RECOVERING = "recovering" -BASELINE_STATUS_LEARNING = "learning" -# ── 状态结构 ────────────────────────────────────────────────────────────────── +# ============================================================================= +# 状态结构 +# ============================================================================= @dataclass class BaselineState: period: int + phase_origin_ts: int template: List[float] status: str clean_seconds: int @@ -160,7 +162,9 @@ class BaselineState: LAST_WRITTEN_UNTIL: Dict[str, int] = {} -# ── VM 读取 ─────────────────────────────────────────────────────────────────── +# ============================================================================= +# VictoriaMetrics 读取 +# ============================================================================= def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: now = datetime.now() @@ -258,7 +262,25 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. return ts_grid, ys_grid -# ── 周期估计 ────────────────────────────────────────────────────────────────── +# ============================================================================= +# 周期估计 +# ============================================================================= + +def moving_average(arr: np.ndarray, window: int) -> np.ndarray: + if window <= 1 or len(arr) < window: + return arr.astype(float) + + window = int(window) + + if window % 2 == 0: + window += 1 + + kernel = np.ones(window, dtype=float) / window + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + return np.convolve(padded, kernel, mode="valid") + def estimate_period_by_fft(ys_arr: np.ndarray) -> float: n = len(ys_arr) @@ -307,7 +329,7 @@ def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: corr = np.correlate(centered, centered, mode="full")[n - 1:] p0 = int(round(init_period)) - left = max(MIN_PERIOD_SECONDS, int(max(2, p0 * 0.7))) + left = max(int(MIN_PERIOD_SECONDS), int(max(2, p0 * 0.7))) right = min(n // 2, int(max(left + 1, p0 * 1.3))) if right <= left: @@ -323,96 +345,252 @@ def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: return float(np.clip(best_lag, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) -def estimate_period(ys_arr: np.ndarray) -> int: +def estimate_period_rough(ys_arr: np.ndarray) -> int: p_fft = estimate_period_by_fft(ys_arr) p_refined = refine_period_by_autocorr(ys_arr, p_fft) period = int(round(p_refined)) - period = max(MIN_PERIOD_SECONDS, min(MAX_PERIOD_SECONDS, period)) + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) return int(period) -# ── 模板构建与预测 ───────────────────────────────────────────────────────────── +# ============================================================================= +# 谷底锚点检测 +# ============================================================================= -def fill_template_nan(template: np.ndarray) -> np.ndarray: - period = len(template) +def find_valley_indices( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + expected_period: int, +) -> List[int]: + n = len(ys_grid) - if period == 0: - return template + if n < max(10, expected_period * 2): + return [] + + period = max(3, int(expected_period)) + + smooth_window = max(3, int(round(period * 0.08))) + smooth_window = min(smooth_window, 21) + + ys_smooth = moving_average(ys_grid, smooth_window) + threshold = float(np.percentile(ys_smooth, VALLEY_QUANTILE)) + + candidates = [] + + for i in range(1, n - 1): + if ( + ys_smooth[i] <= ys_smooth[i - 1] + and ys_smooth[i] < ys_smooth[i + 1] + and ys_smooth[i] <= threshold + ): + candidates.append(i) + + if len(candidates) < MIN_FULL_CYCLES_FOR_TEMPLATE: + candidates = [] + + for i in range(1, n - 1): + if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1]: + candidates.append(i) + + if not candidates: + return [] - idx = np.arange(period) - valid = np.isfinite(template) + min_distance = max(2, int(round(period * 0.55))) + selected = [] - if not np.any(valid): - return np.zeros(period, dtype=float) + for idx in candidates: + if not selected: + selected.append(idx) + continue + + if idx - selected[-1] >= min_distance: + selected.append(idx) + continue + + if ys_smooth[idx] < ys_smooth[selected[-1]]: + selected[-1] = idx - if np.all(valid): - return template + if len(selected) < 2: + return selected - x_valid = idx[valid] - y_valid = template[valid] + cleaned = [selected[0]] + + for idx in selected[1:]: + diff = int(ts_grid[idx] - ts_grid[cleaned[-1]]) + + if int(period * 0.55) <= diff <= int(period * 1.60): + cleaned.append(idx) + continue - # 环形插值，处理 phase 0 附近缺口 - x_ext = np.concatenate([x_valid - period, x_valid, x_valid + period]) - y_ext = np.concatenate([y_valid, y_valid, y_valid]) + if diff < int(period * 0.55): + if ys_smooth[idx] < ys_smooth[cleaned[-1]]: + cleaned[-1] = idx + continue - filled = np.interp(idx, x_ext, y_ext) + cleaned.append(idx) - return filled.astype(float) + return cleaned -def build_phase_template( +def detect_period_and_valleys( + ts_grid: np.ndarray, + ys_grid: np.ndarray, +) -> Tuple[int, List[int]]: + rough = estimate_period_rough(ys_grid) + valleys = find_valley_indices(ts_grid, ys_grid, rough) + + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys]) + good = diffs[(diffs >= rough * 0.55) & (diffs <= rough * 1.60)] + + if len(good) > 0: + period = int(round(float(np.median(good)))) + else: + period = rough + else: + period = rough + + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) + + return int(period), valleys + + +# ============================================================================= +# 相位对齐模板构建 +# ============================================================================= + +def build_template_from_valleys( ts_grid: np.ndarray, ys_grid: np.ndarray, period: int, + valleys: List[int], max_cycles: int = MAX_CYCLES_FOR_TEMPLATE, - tail_seconds: Optional[int] = None, ) -> Optional[np.ndarray]: - if period <= 1 or len(ys_grid) < period * MIN_FULL_CYCLES_FOR_TEMPLATE: + if period <= 1 or len(valleys) < MIN_FULL_CYCLES_FOR_TEMPLATE + 1: return None - max_seconds = period * max_cycles + pairs = [] + + for a, b in zip(valleys[:-1], valleys[1:]): + cycle_len = float(ts_grid[b] - ts_grid[a]) - if tail_seconds is not None: - max_seconds = min(max_seconds, int(tail_seconds)) + if period * 0.55 <= cycle_len <= period * 1.60: + pairs.append((a, b, cycle_len)) - max_seconds = max(period * MIN_FULL_CYCLES_FOR_TEMPLATE, max_seconds) + if len(pairs) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + pairs = pairs[-max_cycles:] + + phase_grid = np.arange(period, dtype=float) + segments = [] + weights = [] + + for idx, (a, b, cycle_len) in enumerate(pairs): + seg_ts = ts_grid[a:b + 1] + seg_y = ys_grid[a:b + 1] + + if len(seg_y) < 3: + continue - if len(ys_grid) < max_seconds: - start_idx = 0 + x_old = (seg_ts - seg_ts[0]) / cycle_len * period + seg = np.interp(phase_grid, x_old, seg_y) + + segments.append(seg.astype(float)) + + weight = 0.5 + 0.5 * ((idx + 1) / len(pairs)) + weights.append(weight) + + if len(segments) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + arr = np.vstack(segments) + w_arr = np.array(weights, dtype=float) + + template = np.average(arr, axis=0, weights=w_arr) + + return template.astype(float) + + +def build_current_baseline( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + tail_seconds: Optional[int] = None, +) -> Optional[Tuple[int, int, np.ndarray]]: + if len(ys_grid) < MIN_POINTS: + return None + + if tail_seconds is not None and tail_seconds > 0: + cutoff = ts_grid[-1] - int(tail_seconds) + mask = ts_grid >= cutoff + ts_use = ts_grid[mask] + ys_use = ys_grid[mask] else: - start_idx = len(ys_grid) - max_seconds + ts_use = ts_grid + ys_use = ys_grid + + if len(ys_use) < MIN_POINTS: + return None - ts_tail = ts_grid[start_idx:] - ys_tail = ys_grid[start_idx:] + period, valleys = detect_period_and_valleys(ts_use, ys_use) - if len(ys_tail) < period * MIN_FULL_CYCLES_FOR_TEMPLATE: + template = build_template_from_valleys( + ts_grid=ts_use, + ys_grid=ys_use, + period=period, + valleys=valleys, + ) + + if template is None or len(valleys) == 0: return None - sums = np.zeros(period, dtype=float) - weights = np.zeros(period, dtype=float) + phase_origin_ts = int(round(float(ts_use[valleys[-1]]))) - total = len(ys_tail) + return int(period), phase_origin_ts, template - for i, (t, y) in enumerate(zip(ts_tail, ys_tail)): - phase = int(t) % period - # 越近的数据权重越高 - recency = (i + 1) / total - weight = 0.3 + 0.7 * recency +# ============================================================================= +# 模板预测 +# ============================================================================= - sums[phase] += float(y) * weight - weights[phase] += weight +def circular_template_value(template: np.ndarray, phase: float) -> float: + period = len(template) - template = np.full(period, np.nan, dtype=float) + if period == 0: + return 0.0 + + phase = float(phase) % period - valid = weights > 0 - template[valid] = sums[valid] / weights[valid] + i0 = int(math.floor(phase)) % period + i1 = (i0 + 1) % period - template = fill_template_nan(template) + frac = phase - math.floor(phase) - return template + return float((1.0 - frac) * template[i0] + frac * template[i1]) + + +def predict_with_origin( + state: BaselineState, + ts_list: List[int], + phase_origin_ts: Optional[int] = None, +) -> np.ndarray: + template = np.array(state.template, dtype=float) + period = int(state.period) + + if period <= 1 or len(template) != period: + return np.zeros(len(ts_list), dtype=float) + + origin = int(state.phase_origin_ts if phase_origin_ts is None else phase_origin_ts) + + values = [] + + for ts in ts_list: + phase = (int(ts) - origin) % period + values.append(circular_template_value(template, phase)) + + return np.array(values, dtype=float) def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: @@ -433,6 +611,38 @@ def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: return np.interp(new_x, old_x_ext, old_y_ext).astype(float) +def align_new_template_to_old( + old_template: np.ndarray, + new_template: np.ndarray, +) -> np.ndarray: + if len(old_template) != len(new_template): + old_template = resample_template(old_template, len(new_template)) + + period = len(new_template) + + if period <= 2: + return new_template.astype(float) + + max_shift = max(1, int(round(period * 0.10))) + + old_norm = old_template - np.mean(old_template) + + best_score = None + best_template = new_template + + for shift in range(-max_shift, max_shift + 1): + shifted = np.roll(new_template, shift) + shifted_norm = shifted - np.mean(shifted) + + score = float(np.dot(old_norm, shifted_norm)) + + if best_score is None or score > best_score: + best_score = score + best_template = shifted + + return best_template.astype(float) + + def merge_template( old_template: np.ndarray, new_template: np.ndarray, @@ -443,37 +653,64 @@ def merge_template( if len(old_template) != len(new_template): old_template = resample_template(old_template, len(new_template)) - return ((1.0 - alpha) * old_template + alpha * new_template).astype(float) - - -def predict_by_state(state: BaselineState, ts_list: List[int]) -> np.ndarray: - template = np.array(state.template, dtype=float) - period = int(state.period) - - if period <= 1 or len(template) != period: - return np.zeros(len(ts_list), dtype=float) + new_template = align_new_template_to_old(old_template, new_template) - values = [] + merged = (1.0 - alpha) * old_template + alpha * new_template - for ts in ts_list: - phase = int(ts) % period - values.append(float(template[phase])) + return merged.astype(float) - return np.array(values, dtype=float) +# ============================================================================= +# 异常检测 +# ============================================================================= -def calc_threshold(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> np.ndarray: +def calc_threshold( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> np.ndarray: return np.maximum(abs_threshold, np.abs(pred) * rel_threshold) -def calc_bounds(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> Tuple[np.ndarray, np.ndarray]: +def calc_bounds( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[np.ndarray, np.ndarray]: threshold = calc_threshold(pred, abs_threshold, rel_threshold) + lower = pred - threshold upper = pred + threshold + return lower, upper -# ── 异常检测与状态更新 ──────────────────────────────────────────────────────── +def find_best_phase_origin_for_recent( + state: BaselineState, + ts_recent: List[int], + actual: np.ndarray, +) -> Tuple[int, np.ndarray, float]: + period = int(state.period) + base_origin = int(state.phase_origin_ts) + + max_shift = max(1, int(round(period * PHASE_SEARCH_RATIO))) + + best_origin = base_origin + best_pred = predict_with_origin(state, ts_recent, base_origin) + best_mae = float(np.mean(np.abs(actual - best_pred))) + + for shift in range(-max_shift, max_shift + 1): + origin = base_origin + shift + pred = predict_with_origin(state, ts_recent, origin) + mae = float(np.mean(np.abs(actual - pred))) + + if mae < best_mae: + best_mae = mae + best_origin = origin + best_pred = pred + + return best_origin, best_pred, best_mae + def detect_anomaly( state: BaselineState, @@ -481,14 +718,19 @@ def detect_anomaly( ys_grid: np.ndarray, abs_threshold: float, rel_threshold: float, -) -> Tuple[bool, float, float, float]: +) -> Tuple[bool, float, float, float, int]: if len(ys_grid) < DETECT_WINDOW_SECONDS: - return False, 0.0, 0.0, 0.0 + return False, 0.0, 0.0, 0.0, int(state.phase_origin_ts) ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) - pred = predict_by_state(state, ts_recent) + best_origin, pred, _ = find_best_phase_origin_for_recent( + state=state, + ts_recent=ts_recent, + actual=actual, + ) + threshold = calc_threshold(pred, abs_threshold, rel_threshold) abs_err = np.abs(actual - pred) @@ -500,56 +742,28 @@ def detect_anomaly( is_anomaly = outside_ratio >= OUTSIDE_RATIO_THRESHOLD - return is_anomaly, outside_ratio, mean_abs_err, mean_rel_err - - -def is_recovered( - state: BaselineState, - ts_grid: np.ndarray, - ys_grid: np.ndarray, - abs_threshold: float, - rel_threshold: float, -) -> Tuple[bool, float]: - if len(ys_grid) < DETECT_WINDOW_SECONDS: - return False, 0.0 - - ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() - actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) - - pred = predict_by_state(state, ts_recent) - threshold = calc_threshold(pred, abs_threshold, rel_threshold) - - abs_err = np.abs(actual - pred) - inside = abs_err <= threshold - - inside_ratio = float(np.mean(inside)) + return is_anomaly, outside_ratio, mean_abs_err, mean_rel_err, int(best_origin) - return inside_ratio >= RECOVERY_INSIDE_RATIO_THRESHOLD, inside_ratio +# ============================================================================= +# 健康基线状态管理 +# ============================================================================= def create_initial_state( ts_grid: np.ndarray, ys_grid: np.ndarray, now_sec: int, ) -> Optional[BaselineState]: - if len(ys_grid) < MIN_POINTS: - return None - - period = estimate_period(ys_grid) - - template = build_phase_template( - ts_grid=ts_grid, - ys_grid=ys_grid, - period=period, - max_cycles=MAX_CYCLES_FOR_TEMPLATE, - tail_seconds=period * MAX_CYCLES_FOR_TEMPLATE, - ) + baseline = build_current_baseline(ts_grid, ys_grid) - if template is None: + if baseline is None: return None + period, phase_origin_ts, template = baseline + return BaselineState( period=int(period), + phase_origin_ts=int(phase_origin_ts), template=template.astype(float).tolist(), status=BASELINE_STATUS_HEALTHY, clean_seconds=int(period * MAX_CYCLES_FOR_TEMPLATE), @@ -578,19 +792,23 @@ def maybe_update_state( return None, False, 0.0, 0.0, 0.0 BASELINE_STATES[key] = state + logger.info( - "初始化健康模板 key=%s period=%ss clean_seconds=%ss", + "初始化健康模板 key=%s period=%ss origin=%s clean=%ss", key, state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), state.clean_seconds, ) + return state, False, 0.0, 0.0, 0.0 elapsed = max(1, now_sec - int(state.last_seen_ts)) elapsed = min(elapsed, POLL_INTERVAL * 2) + state.last_seen_ts = now_sec - is_anom, outside_ratio, mean_abs_err, mean_rel_err = detect_anomaly( + is_anom, outside_ratio, mean_abs_err, mean_rel_err, best_origin = detect_anomaly( state=state, ts_grid=ts_grid, ys_grid=ys_grid, @@ -602,6 +820,8 @@ def maybe_update_state( state.status = BASELINE_STATUS_ANOMALY state.clean_seconds = 0 + BASELINE_STATES[key] = state + logger.warning( "检测到异常，冻结模板 key=%s outside_ratio=%.2f mean_abs_err=%.2f mean_rel_err=%.2f", key, @@ -610,47 +830,39 @@ def maybe_update_state( mean_rel_err, ) - BASELINE_STATES[key] = state return state, True, outside_ratio, mean_abs_err, mean_rel_err - recovered, inside_ratio = is_recovered( - state=state, - ts_grid=ts_grid, - ys_grid=ys_grid, - abs_threshold=abs_threshold, - rel_threshold=rel_threshold, - ) + old_origin = int(state.phase_origin_ts) + state.phase_origin_ts = int(best_origin) + + if abs(state.phase_origin_ts - old_origin) >= 1: + logger.debug( + "相位校正 key=%s origin %s -> %s", + key, + datetime.fromtimestamp(old_origin).strftime("%H:%M:%S"), + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + ) if state.status == BASELINE_STATUS_ANOMALY: - if recovered: - state.status = BASELINE_STATUS_RECOVERING - state.clean_seconds = elapsed - logger.info( - "异常开始恢复 key=%s inside_ratio=%.2f clean_seconds=%ss", - key, - inside_ratio, - state.clean_seconds, - ) - else: - state.clean_seconds = 0 - BASELINE_STATES[key] = state - return state, True, outside_ratio, mean_abs_err, mean_rel_err + state.status = BASELINE_STATUS_RECOVERING + state.clean_seconds = elapsed - elif state.status == BASELINE_STATUS_RECOVERING: - if recovered: - state.clean_seconds += elapsed - else: - state.status = BASELINE_STATUS_ANOMALY - state.clean_seconds = 0 - BASELINE_STATES[key] = state - return state, True, outside_ratio, mean_abs_err, mean_rel_err + BASELINE_STATES[key] = state + logger.info( + "异常开始恢复 key=%s clean_seconds=%ss", + key, + state.clean_seconds, + ) + + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + if state.status == BASELINE_STATUS_RECOVERING: + state.clean_seconds += elapsed else: state.status = BASELINE_STATUS_HEALTHY state.clean_seconds += elapsed - # 故障恢复后，不要立刻学习。 - # 必须至少连续正常：max(RECOVERY_MIN_SECONDS, 3 个周期) min_clean_for_update = max( RECOVERY_MIN_SECONDS, int(state.period) * MIN_FULL_CYCLES_FOR_TEMPLATE, @@ -660,30 +872,26 @@ def maybe_update_state( BASELINE_STATES[key] = state return state, False, outside_ratio, mean_abs_err, mean_rel_err - # 只使用最近 clean_seconds 这段连续正常数据来更新模板，避免历史故障污染。 - new_period = estimate_period(ys_grid) tail_seconds = min( int(state.clean_seconds), - int(new_period) * MAX_CYCLES_FOR_TEMPLATE, + int(state.period) * MAX_CYCLES_FOR_TEMPLATE, ) - new_template = build_phase_template( + baseline = build_current_baseline( ts_grid=ts_grid, ys_grid=ys_grid, - period=new_period, - max_cycles=MAX_CYCLES_FOR_TEMPLATE, tail_seconds=tail_seconds, ) - if new_template is None: + if baseline is None: BASELINE_STATES[key] = state return state, False, outside_ratio, mean_abs_err, mean_rel_err + new_period, new_origin, new_template = baseline old_template = np.array(state.template, dtype=float) if state.status == BASELINE_STATUS_RECOVERING: alpha = RECOVERY_EMA_ALPHA - state.status = BASELINE_STATUS_HEALTHY else: alpha = HEALTHY_EMA_ALPHA @@ -694,18 +902,25 @@ def maybe_update_state( ) state.period = int(new_period) + state.phase_origin_ts = int(new_origin) state.template = merged.astype(float).tolist() + state.status = BASELINE_STATUS_HEALTHY state.last_update_ts = now_sec - state.y_min = float(np.min(ys_grid[-tail_seconds:])) - state.y_max = float(np.max(ys_grid[-tail_seconds:])) + + if tail_seconds > 0 and len(ys_grid) >= tail_seconds: + state.y_min = float(np.min(ys_grid[-tail_seconds:])) + state.y_max = float(np.max(ys_grid[-tail_seconds:])) + else: + state.y_min = float(np.min(ys_grid)) + state.y_max = float(np.max(ys_grid)) BASELINE_STATES[key] = state logger.info( - "更新健康模板 key=%s period=%ss status=%s clean_seconds=%ss alpha=%.2f", + "更新健康模板 key=%s period=%ss origin=%s clean=%ss alpha=%.2f", key, state.period, - state.status, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), state.clean_seconds, alpha, ) @@ -713,7 +928,9 @@ def maybe_update_state( return state, False, outside_ratio, mean_abs_err, mean_rel_err -# ── Prometheus 格式写入 ─────────────────────────────────────────────────────── +# ============================================================================= +# Prometheus Exposition 写入 +# ============================================================================= def prom_escape_label_value(value: str) -> str: return ( @@ -731,8 +948,7 @@ def labels_to_str(labels: Dict[str, str]) -> str: parts = [] for k in sorted(labels.keys()): - v = prom_escape_label_value(labels[k]) - parts.append(f'{k}="{v}"') + parts.append(f'{k}="{prom_escape_label_value(labels[k])}"') return "{" + ",".join(parts) + "}" @@ -771,11 +987,14 @@ def write_series( resp = requests.post( f"{VM_URL}/api/v1/import/prometheus", data=payload.encode("utf-8"), - headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + headers={ + "Content-Type": "text/plain; version=0.0.4; charset=utf-8", + }, timeout=10, ) resp.raise_for_status() return True + except requests.RequestException as e: logger.error("写入数据失败 metric=%s: %s", metric_name, e) return False @@ -851,28 +1070,37 @@ def write_prediction_bundle( return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 -# ── 标签解析 ────────────────────────────────────────────────────────────────── +# ============================================================================= +# 标签解析 +# ============================================================================= _LABEL_PATTERN = re.compile( r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' ) -def _parse_labels(query: str) -> Dict[str, str]: +def parse_labels_from_query(query: str) -> Dict[str, str]: labels = {} if "{" not in query or "}" not in query: return labels try: - label_part = query[query.index("{") + 1: query.rindex("}")] + label_part = query[query.index("{") + 1:query.rindex("}")] except Exception: return labels for match in _LABEL_PATTERN.finditer(label_part): key = match.group(1) value = match.group(2) - value = value.replace('\\"', '"').replace("\\n", "\n").replace("\\\\", "\\") + + value = ( + value + .replace('\\"', '"') + .replace("\\n", "\n") + .replace("\\\\", "\\") + ) + labels[key] = value return labels @@ -892,9 +1120,11 @@ def series_key(metric_name: str, labels: Dict[str, str]) -> str: return metric_name + labels_to_str(labels) -# ── 状态持久化 ──────────────────────────────────────────────────────────────── +# ============================================================================= +# 状态持久化 +# ============================================================================= -def load_state(): +def load_state() -> None: global BASELINE_STATES if not os.path.exists(STATE_FILE): @@ -907,17 +1137,36 @@ def load_state(): states = {} for key, value in raw.get("baseline_states", {}).items(): + required_fields = { + "period", + "phase_origin_ts", + "template", + "status", + "clean_seconds", + "last_update_ts", + "last_seen_ts", + "y_min", + "y_max", + } + + if not required_fields.issubset(set(value.keys())): + continue + states[key] = BaselineState(**value) BASELINE_STATES = states - logger.info("已加载预测状态文件 %s，状态数量=%d", STATE_FILE, len(BASELINE_STATES)) + logger.info( + "已加载预测状态文件 %s，状态数量=%d", + STATE_FILE, + len(BASELINE_STATES), + ) except Exception as e: logger.warning("加载预测状态文件失败，将重新学习: %s", e) -def save_state(): +def save_state() -> None: try: raw = { "baseline_states": { @@ -937,9 +1186,11 @@ def save_state(): logger.warning("保存预测状态文件失败: %s", e) -# ── 主逻辑 ──────────────────────────────────────────────────────────────────── +# ============================================================================= +# 主流程 +# ============================================================================= -def run_once(): +def run_once() -> None: now_str = datetime.now().strftime("%H:%M:%S") for target in PREDICT_TARGETS: @@ -952,16 +1203,26 @@ def run_once(): ts, ys = fetch_history(query) if len(ys) < MIN_POINTS: - logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + logger.info( + "[%s] %s 数据不足（%d 点），跳过", + now_str, + query, + len(ys), + ) continue ts_grid, ys_grid = normalize_history(ts, ys) if len(ys_grid) < MIN_POINTS: - logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid)) + logger.info( + "[%s] %s 清洗后数据不足（%d 点），跳过", + now_str, + query, + len(ys_grid), + ) continue - base_labels = _parse_labels(query) + base_labels = parse_labels_from_query(query) write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) key = series_key(pred_metric, write_labels) @@ -975,7 +1236,11 @@ def run_once(): ) if state is None: - logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + logger.info( + "[%s] %s 暂无可用健康模板，等待学习", + now_str, + query, + ) continue now_sec = int(time.time()) @@ -989,7 +1254,7 @@ def run_once(): for i in range(WRITE_HORIZON_SECONDS) ] - pred_values = predict_by_state(state, ts_future) + pred_values = predict_with_origin(state, ts_future) lower_values, upper_values = calc_bounds( pred=pred_values, @@ -1012,21 +1277,28 @@ def run_once(): ) if not ok: + logger.error( + "[%s] %s 写入预测数据失败", + now_str, + query, + ) continue LAST_WRITTEN_UNTIL[key] = int(max(ts_future)) future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") logger.info( - "[%s] %-40s → %-35s status=%s anomaly=%s period=%ss clean=%ss 写入 %d 点，预测区间 %s ~ %s", + "[%s] %-40s → %-35s status=%s anomaly=%s period=%ss origin=%s clean=%ss 写入 %d 点，预测区间 %s ~ %s", now_str, query, pred_metric, state.status, is_anomaly, state.period, + origin_str, state.clean_seconds, len(ts_future), future_start, @@ -1036,17 +1308,18 @@ def run_once(): save_state() -def main(): +def main() -> None: load_state() logger.info( - "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s", + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s", VM_URL, HISTORY_MINUTES, HORIZON_SECONDS, WRITE_HORIZON_SECONDS, POLL_INTERVAL, STATE_FILE, + EXTRA_PREDICT_LABELS["forecast"], ) while True: @@ -1055,4 +1328,6 @@ def main(): if __name__ == "__main__": - main() \ No newline at end of file + main() + + \ No newline at end of file From 76e536eff9beb9e3f9db7453bbeb6f1c7844c198 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 13:55:08 +0800 Subject: [PATCH 24/36] fix --- ai/predict_v3_single_scene.py | 316 ++++++++++++---------------------- 1 file changed, 110 insertions(+), 206 deletions(-) diff --git a/ai/predict_v3_single_scene.py b/ai/predict_v3_single_scene.py index fc07f4f..2cde8b8 100755 --- a/ai/predict_v3_single_scene.py +++ b/ai/predict_v3_single_scene.py @@ -1,22 +1,12 @@ # -*- coding: utf-8 -*- """ -ProtoForge Predictor v8 - -功能： -1. 从 VictoriaMetrics 拉取历史数据。 -2. 对 CNC 周期型指标进行相位对齐预测。 -3. 使用“谷底锚点”对齐周期，减少上升沿/下降沿相位偏差。 -4. 每轮只写入未来 min(HORIZON_SECONDS, POLL_INTERVAL) 秒，避免预测窗口重叠。 -5. 检测异常后冻结健康模板，不把故障数据学进去。 -6. 故障恢复后等待稳定一段时间，再恢复模板更新。 -7. 写入： - - xxx_predicted - - xxx_predicted_upper - - xxx_predicted_lower - - xxx_anomaly - - xxx_anomaly_outside_ratio - - xxx_anomaly_mean_abs_error - - xxx_anomaly_mean_rel_error +ProtoForge Predictor v9 + +修复重点： +1. 预测时间轴改为锚定最后一个真实数据点 last_real_ts，而不是锚定 time.time()。 +2. 不再使用 LAST_WRITTEN_UNTIL 把预测不断推向更远未来，避免 Grafana 里预测线相对真实线出现延迟/错位。 +3. 如果真实数据时间戳没有推进，则跳过本轮预测写入，避免重复写同一段未来时间造成毛刺。 +4. 保留：相位对齐、健康模板冻结、故障期不学习、恢复后再学习、预测上下界、异常指标。 """ import json @@ -50,14 +40,13 @@ # ============================================================================= VM_URL = "http://localhost:8428" - -STATE_FILE = "/tmp/protoforge_predictor_state_v8.json" +STATE_FILE = "/tmp/protoforge_predictor_state_v9.json" HISTORY_MINUTES = 30 HORIZON_SECONDS = 120 POLL_INTERVAL = 30 -# 实际写入窗口不要大于轮询间隔，否则多轮预测会重叠。 +# 实际每轮写入的预测长度。不要大于 POLL_INTERVAL，否则容易出现预测窗口重叠。 WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) QUERY_STEP = "1s" @@ -76,14 +65,23 @@ RECOVERY_EMA_ALPHA = 0.30 OUTSIDE_RATIO_THRESHOLD = 0.60 -RECOVERY_INSIDE_RATIO_THRESHOLD = 0.80 - PHASE_SEARCH_RATIO = 0.15 VALLEY_QUANTILE = 45 +# 关键修复：预测时间轴锚定真实数据最后一个点。 +# True：预测从 last_real_ts + 1 开始，适合 Grafana 与真实曲线对齐展示。 +# False：预测从当前系统时间 + 1 开始，适合只看纯未来预测，但容易与有采集延迟的真实数据错位。 +ALIGN_PREDICTION_TO_LAST_REAL_TS = True + +# 如果 last_real_ts 距离当前系统时间太久，说明采集链路可能断了，跳过预测，避免用陈旧数据继续画未来线。 +MAX_DATA_LAG_SECONDS = 180 + +# 真实数据至少推进多少秒，才写入新预测，避免同一段未来时间被反复写入。 +MIN_REAL_ADVANCE_SECONDS = 1 + # ============================================================================= -# 预测指标配置 +# 指标配置 # ============================================================================= PREDICT_TARGETS = [ @@ -132,7 +130,7 @@ ] EXTRA_PREDICT_LABELS = { - "forecast": "phase_aligned_health_v8", + "forecast": "phase_aligned_health_v9", "source": "protoforge", } @@ -159,7 +157,10 @@ class BaselineState: BASELINE_STATES: Dict[str, BaselineState] = {} -LAST_WRITTEN_UNTIL: Dict[str, int] = {} + +# 记录每条序列最后一次使用的真实数据时间戳，而不是预测写到哪里。 +# 这样不会把预测不断推向更远的未来。 +LAST_REAL_TS_WRITTEN: Dict[str, int] = {} # ============================================================================= @@ -243,7 +244,6 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. return np.array([]), np.array([]) sorted_items = sorted(data.items(), key=lambda x: x[0]) - ts_clean = np.array([x[0] for x in sorted_items], dtype=float) ys_clean = np.array([x[1] for x in sorted_items], dtype=float) @@ -263,7 +263,7 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. # ============================================================================= -# 周期估计 +# 周期估计与谷底检测 # ============================================================================= def moving_average(arr: np.ndarray, window: int) -> np.ndarray: @@ -355,10 +355,6 @@ def estimate_period_rough(ys_arr: np.ndarray) -> int: return int(period) -# ============================================================================= -# 谷底锚点检测 -# ============================================================================= - def find_valley_indices( ts_grid: np.ndarray, ys_grid: np.ndarray, @@ -370,7 +366,6 @@ def find_valley_indices( return [] period = max(3, int(expected_period)) - smooth_window = max(3, int(round(period * 0.08))) smooth_window = min(smooth_window, 21) @@ -389,7 +384,6 @@ def find_valley_indices( if len(candidates) < MIN_FULL_CYCLES_FOR_TEMPLATE: candidates = [] - for i in range(1, n - 1): if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1]: candidates.append(i) @@ -458,7 +452,7 @@ def detect_period_and_valleys( # ============================================================================= -# 相位对齐模板构建 +# 相位对齐模板 # ============================================================================= def build_template_from_valleys( @@ -552,7 +546,7 @@ def build_current_baseline( # ============================================================================= -# 模板预测 +# 预测与模板合并 # ============================================================================= def circular_template_value(template: np.ndarray, phase: float) -> float: @@ -562,10 +556,8 @@ def circular_template_value(template: np.ndarray, phase: float) -> float: return 0.0 phase = float(phase) % period - i0 = int(math.floor(phase)) % period i1 = (i0 + 1) % period - frac = phase - math.floor(phase) return float((1.0 - frac) * template[i0] + frac * template[i1]) @@ -583,7 +575,6 @@ def predict_with_origin( return np.zeros(len(ts_list), dtype=float) origin = int(state.phase_origin_ts if phase_origin_ts is None else phase_origin_ts) - values = [] for ts in ts_list: @@ -611,10 +602,7 @@ def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: return np.interp(new_x, old_x_ext, old_y_ext).astype(float) -def align_new_template_to_old( - old_template: np.ndarray, - new_template: np.ndarray, -) -> np.ndarray: +def align_new_template_to_old(old_template: np.ndarray, new_template: np.ndarray) -> np.ndarray: if len(old_template) != len(new_template): old_template = resample_template(old_template, len(new_template)) @@ -624,7 +612,6 @@ def align_new_template_to_old( return new_template.astype(float) max_shift = max(1, int(round(period * 0.10))) - old_norm = old_template - np.mean(old_template) best_score = None @@ -633,7 +620,6 @@ def align_new_template_to_old( for shift in range(-max_shift, max_shift + 1): shifted = np.roll(new_template, shift) shifted_norm = shifted - np.mean(shifted) - score = float(np.dot(old_norm, shifted_norm)) if best_score is None or score > best_score: @@ -643,18 +629,13 @@ def align_new_template_to_old( return best_template.astype(float) -def merge_template( - old_template: np.ndarray, - new_template: np.ndarray, - alpha: float, -) -> np.ndarray: +def merge_template(old_template: np.ndarray, new_template: np.ndarray, alpha: float) -> np.ndarray: alpha = float(np.clip(alpha, 0.0, 1.0)) if len(old_template) != len(new_template): old_template = resample_template(old_template, len(new_template)) new_template = align_new_template_to_old(old_template, new_template) - merged = (1.0 - alpha) * old_template + alpha * new_template return merged.astype(float) @@ -664,11 +645,7 @@ def merge_template( # 异常检测 # ============================================================================= -def calc_threshold( - pred: np.ndarray, - abs_threshold: float, - rel_threshold: float, -) -> np.ndarray: +def calc_threshold(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> np.ndarray: return np.maximum(abs_threshold, np.abs(pred) * rel_threshold) @@ -678,11 +655,7 @@ def calc_bounds( rel_threshold: float, ) -> Tuple[np.ndarray, np.ndarray]: threshold = calc_threshold(pred, abs_threshold, rel_threshold) - - lower = pred - threshold - upper = pred + threshold - - return lower, upper + return pred - threshold, pred + threshold def find_best_phase_origin_for_recent( @@ -692,7 +665,6 @@ def find_best_phase_origin_for_recent( ) -> Tuple[int, np.ndarray, float]: period = int(state.period) base_origin = int(state.phase_origin_ts) - max_shift = max(1, int(round(period * PHASE_SEARCH_RATIO))) best_origin = base_origin @@ -732,14 +704,12 @@ def detect_anomaly( ) threshold = calc_threshold(pred, abs_threshold, rel_threshold) - abs_err = np.abs(actual - pred) outside = abs_err > threshold outside_ratio = float(np.mean(outside)) mean_abs_err = float(np.mean(abs_err)) mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1.0))) - is_anomaly = outside_ratio >= OUTSIDE_RATIO_THRESHOLD return is_anomaly, outside_ratio, mean_abs_err, mean_rel_err, int(best_origin) @@ -749,11 +719,7 @@ def detect_anomaly( # 健康基线状态管理 # ============================================================================= -def create_initial_state( - ts_grid: np.ndarray, - ys_grid: np.ndarray, - now_sec: int, -) -> Optional[BaselineState]: +def create_initial_state(ts_grid: np.ndarray, ys_grid: np.ndarray, now_sec: int) -> Optional[BaselineState]: baseline = build_current_baseline(ts_grid, ys_grid) if baseline is None: @@ -782,7 +748,6 @@ def maybe_update_state( rel_threshold: float, ) -> Tuple[Optional[BaselineState], bool, float, float, float]: now_sec = int(time.time()) - state = BASELINE_STATES.get(key) if state is None: @@ -805,7 +770,6 @@ def maybe_update_state( elapsed = max(1, now_sec - int(state.last_seen_ts)) elapsed = min(elapsed, POLL_INTERVAL * 2) - state.last_seen_ts = now_sec is_anom, outside_ratio, mean_abs_err, mean_rel_err, best_origin = detect_anomaly( @@ -819,7 +783,6 @@ def maybe_update_state( if is_anom: state.status = BASELINE_STATUS_ANOMALY state.clean_seconds = 0 - BASELINE_STATES[key] = state logger.warning( @@ -846,15 +809,9 @@ def maybe_update_state( if state.status == BASELINE_STATUS_ANOMALY: state.status = BASELINE_STATUS_RECOVERING state.clean_seconds = elapsed - BASELINE_STATES[key] = state - logger.info( - "异常开始恢复 key=%s clean_seconds=%ss", - key, - state.clean_seconds, - ) - + logger.info("异常开始恢复 key=%s clean_seconds=%ss", key, state.clean_seconds) return state, False, outside_ratio, mean_abs_err, mean_rel_err if state.status == BASELINE_STATUS_RECOVERING: @@ -877,11 +834,7 @@ def maybe_update_state( int(state.period) * MAX_CYCLES_FOR_TEMPLATE, ) - baseline = build_current_baseline( - ts_grid=ts_grid, - ys_grid=ys_grid, - tail_seconds=tail_seconds, - ) + baseline = build_current_baseline(ts_grid=ts_grid, ys_grid=ys_grid, tail_seconds=tail_seconds) if baseline is None: BASELINE_STATES[key] = state @@ -889,17 +842,9 @@ def maybe_update_state( new_period, new_origin, new_template = baseline old_template = np.array(state.template, dtype=float) + alpha = RECOVERY_EMA_ALPHA if state.status == BASELINE_STATUS_RECOVERING else HEALTHY_EMA_ALPHA - if state.status == BASELINE_STATUS_RECOVERING: - alpha = RECOVERY_EMA_ALPHA - else: - alpha = HEALTHY_EMA_ALPHA - - merged = merge_template( - old_template=old_template, - new_template=new_template, - alpha=alpha, - ) + merged = merge_template(old_template=old_template, new_template=new_template, alpha=alpha) state.period = int(new_period) state.phase_origin_ts = int(new_origin) @@ -933,12 +878,7 @@ def maybe_update_state( # ============================================================================= def prom_escape_label_value(value: str) -> str: - return ( - str(value) - .replace("\\", "\\\\") - .replace("\n", "\\n") - .replace('"', '\\"') - ) + return str(value).replace("\\", "\\\\").replace("\n", "\\n").replace('"', '\\"') def labels_to_str(labels: Dict[str, str]) -> str: @@ -975,8 +915,7 @@ def write_series( if not math.isfinite(ts_sec) or not math.isfinite(val): continue - ts_ms = ts_sec * 1000 - lines.append(f"{metric_name}{label_str} {val:.6f} {ts_ms}") + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_sec * 1000}") if not lines: return False @@ -987,9 +926,7 @@ def write_series( resp = requests.post( f"{VM_URL}/api/v1/import/prometheus", data=payload.encode("utf-8"), - headers={ - "Content-Type": "text/plain; version=0.0.4; charset=utf-8", - }, + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, timeout=10, ) resp.raise_for_status() @@ -1012,60 +949,19 @@ def write_prediction_bundle( outside_ratio: float, mean_abs_err: float, mean_rel_err: float, + event_ts: int, ) -> bool: - ok1 = write_series( - metric_name=pred_metric, - labels=labels, - ts_list=ts_future, - values=pred_values.astype(float).tolist(), - ) - - ok2 = write_series( - metric_name=f"{pred_metric}_lower", - labels=labels, - ts_list=ts_future, - values=lower_values.astype(float).tolist(), - ) - - ok3 = write_series( - metric_name=f"{pred_metric}_upper", - labels=labels, - ts_list=ts_future, - values=upper_values.astype(float).tolist(), - ) - - now_sec = int(time.time()) + ok1 = write_series(pred_metric, labels, ts_future, pred_values.astype(float).tolist()) + ok2 = write_series(f"{pred_metric}_lower", labels, ts_future, lower_values.astype(float).tolist()) + ok3 = write_series(f"{pred_metric}_upper", labels, ts_future, upper_values.astype(float).tolist()) anomaly_labels = dict(labels) anomaly_labels["type"] = "prediction_deviation" - ok4 = write_series( - metric_name=anomaly_metric, - labels=anomaly_labels, - ts_list=[now_sec], - values=[1.0 if is_anomaly else 0.0], - ) - - ok5 = write_series( - metric_name=f"{anomaly_metric}_outside_ratio", - labels=anomaly_labels, - ts_list=[now_sec], - values=[outside_ratio], - ) - - ok6 = write_series( - metric_name=f"{anomaly_metric}_mean_abs_error", - labels=anomaly_labels, - ts_list=[now_sec], - values=[mean_abs_err], - ) - - ok7 = write_series( - metric_name=f"{anomaly_metric}_mean_rel_error", - labels=anomaly_labels, - ts_list=[now_sec], - values=[mean_rel_err], - ) + ok4 = write_series(anomaly_metric, anomaly_labels, [event_ts], [1.0 if is_anomaly else 0.0]) + ok5 = write_series(f"{anomaly_metric}_outside_ratio", anomaly_labels, [event_ts], [outside_ratio]) + ok6 = write_series(f"{anomaly_metric}_mean_abs_error", anomaly_labels, [event_ts], [mean_abs_err]) + ok7 = write_series(f"{anomaly_metric}_mean_rel_error", anomaly_labels, [event_ts], [mean_rel_err]) return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 @@ -1074,9 +970,7 @@ def write_prediction_bundle( # 标签解析 # ============================================================================= -_LABEL_PATTERN = re.compile( - r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' -) +_LABEL_PATTERN = re.compile(r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*') def parse_labels_from_query(query: str) -> Dict[str, str]: @@ -1093,14 +987,7 @@ def parse_labels_from_query(query: str) -> Dict[str, str]: for match in _LABEL_PATTERN.finditer(label_part): key = match.group(1) value = match.group(2) - - value = ( - value - .replace('\\"', '"') - .replace("\\n", "\n") - .replace("\\\\", "\\") - ) - + value = value.replace('\\"', '"').replace("\\n", "\n").replace("\\\\", "\\") labels[key] = value return labels @@ -1155,12 +1042,7 @@ def load_state() -> None: states[key] = BaselineState(**value) BASELINE_STATES = states - - logger.info( - "已加载预测状态文件 %s，状态数量=%d", - STATE_FILE, - len(BASELINE_STATES), - ) + logger.info("已加载预测状态文件 %s，状态数量=%d", STATE_FILE, len(BASELINE_STATES)) except Exception as e: logger.warning("加载预测状态文件失败，将重新学习: %s", e) @@ -1186,6 +1068,44 @@ def save_state() -> None: logger.warning("保存预测状态文件失败: %s", e) +# ============================================================================= +# 时间轴选择 +# ============================================================================= + +def build_prediction_timestamps(key: str, last_real_ts: int, now_sec: int) -> Optional[List[int]]: + data_lag = now_sec - last_real_ts + + if data_lag > MAX_DATA_LAG_SECONDS: + logger.warning( + "真实数据延迟过大，跳过预测 key=%s data_lag=%ss max=%ss", + key, + data_lag, + MAX_DATA_LAG_SECONDS, + ) + return None + + last_written_real_ts = LAST_REAL_TS_WRITTEN.get(key) + + if last_written_real_ts is not None: + advance = last_real_ts - int(last_written_real_ts) + + if advance < MIN_REAL_ADVANCE_SECONDS: + logger.info( + "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written_real_ts=%s", + key, + last_real_ts, + last_written_real_ts, + ) + return None + + if ALIGN_PREDICTION_TO_LAST_REAL_TS: + base_ts = last_real_ts + else: + base_ts = now_sec + + return [base_ts + i + 1 for i in range(WRITE_HORIZON_SECONDS)] + + # ============================================================================= # 主流程 # ============================================================================= @@ -1203,28 +1123,17 @@ def run_once() -> None: ts, ys = fetch_history(query) if len(ys) < MIN_POINTS: - logger.info( - "[%s] %s 数据不足（%d 点），跳过", - now_str, - query, - len(ys), - ) + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) continue ts_grid, ys_grid = normalize_history(ts, ys) if len(ys_grid) < MIN_POINTS: - logger.info( - "[%s] %s 清洗后数据不足（%d 点），跳过", - now_str, - query, - len(ys_grid), - ) + logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid)) continue base_labels = parse_labels_from_query(query) write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) - key = series_key(pred_metric, write_labels) state, is_anomaly, outside_ratio, mean_abs_err, mean_rel_err = maybe_update_state( @@ -1236,26 +1145,23 @@ def run_once() -> None: ) if state is None: - logger.info( - "[%s] %s 暂无可用健康模板，等待学习", - now_str, - query, - ) + logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) continue now_sec = int(time.time()) - last_until = LAST_WRITTEN_UNTIL.get(key, 0) last_real_ts = int(ts_grid[-1]) + data_lag = now_sec - last_real_ts - base_ts = max(now_sec, last_until, last_real_ts) + ts_future = build_prediction_timestamps( + key=key, + last_real_ts=last_real_ts, + now_sec=now_sec, + ) - ts_future = [ - base_ts + i + 1 - for i in range(WRITE_HORIZON_SECONDS) - ] + if not ts_future: + continue pred_values = predict_with_origin(state, ts_future) - lower_values, upper_values = calc_bounds( pred=pred_values, abs_threshold=abs_threshold, @@ -1274,24 +1180,22 @@ def run_once() -> None: outside_ratio=outside_ratio, mean_abs_err=mean_abs_err, mean_rel_err=mean_rel_err, + event_ts=last_real_ts, ) if not ok: - logger.error( - "[%s] %s 写入预测数据失败", - now_str, - query, - ) + logger.error("[%s] %s 写入预测数据失败", now_str, query) continue - LAST_WRITTEN_UNTIL[key] = int(max(ts_future)) + LAST_REAL_TS_WRITTEN[key] = last_real_ts future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + last_real_str = datetime.fromtimestamp(last_real_ts).strftime("%H:%M:%S") origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") logger.info( - "[%s] %-40s → %-35s status=%s anomaly=%s period=%ss origin=%s clean=%ss 写入 %d 点，预测区间 %s ~ %s", + "[%s] %-40s → %-35s status=%s anomaly=%s period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", now_str, query, pred_metric, @@ -1299,7 +1203,8 @@ def run_once() -> None: is_anomaly, state.period, origin_str, - state.clean_seconds, + last_real_str, + data_lag, len(ts_future), future_start, future_end, @@ -1312,7 +1217,7 @@ def main() -> None: load_state() logger.info( - "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s", + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s align_to_last_real=%s", VM_URL, HISTORY_MINUTES, HORIZON_SECONDS, @@ -1320,6 +1225,7 @@ def main() -> None: POLL_INTERVAL, STATE_FILE, EXTRA_PREDICT_LABELS["forecast"], + ALIGN_PREDICTION_TO_LAST_REAL_TS, ) while True: @@ -1329,5 +1235,3 @@ def main() -> None: if __name__ == "__main__": main() - - \ No newline at end of file From f5e7b2a27320066b5eb9157f297d6651adcd36c8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 14:05:54 +0800 Subject: [PATCH 25/36] fix --- ai/predict_v3_single_scene.py | 520 +++++++++++++++++++++++++--------- 1 file changed, 385 insertions(+), 135 deletions(-) diff --git a/ai/predict_v3_single_scene.py b/ai/predict_v3_single_scene.py index 2cde8b8..d212d2d 100755 --- a/ai/predict_v3_single_scene.py +++ b/ai/predict_v3_single_scene.py @@ -1,12 +1,14 @@ # -*- coding: utf-8 -*- """ -ProtoForge Predictor v9 +ProtoForge Predictor v10 修复重点： -1. 预测时间轴改为锚定最后一个真实数据点 last_real_ts，而不是锚定 time.time()。 -2. 不再使用 LAST_WRITTEN_UNTIL 把预测不断推向更远未来，避免 Grafana 里预测线相对真实线出现延迟/错位。 -3. 如果真实数据时间戳没有推进，则跳过本轮预测写入，避免重复写同一段未来时间造成毛刺。 -4. 保留：相位对齐、健康模板冻结、故障期不学习、恢复后再学习、预测上下界、异常指标。 +1. 修复 lag=0 但预测线仍然相位漂移的问题。 +2. 在谷底相位对齐基础上，增加 phase-lock 相位锁定。 +3. 每轮使用最近 1~2 个周期真实数据，搜索最佳 period + phase_origin。 +4. 预测起点仍然锚定最后一个真实点 last_real_ts，避免写入延迟。 +5. 保留健康模板冻结逻辑：异常期间不学习故障数据。 +6. 保留预测上下界和异常指标。 """ import json @@ -40,13 +42,12 @@ # ============================================================================= VM_URL = "http://localhost:8428" -STATE_FILE = "/tmp/protoforge_predictor_state_v9.json" +STATE_FILE = "/tmp/protoforge_predictor_state_v10.json" HISTORY_MINUTES = 30 HORIZON_SECONDS = 120 POLL_INTERVAL = 30 -# 实际每轮写入的预测长度。不要大于 POLL_INTERVAL，否则容易出现预测窗口重叠。 WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) QUERY_STEP = "1s" @@ -58,26 +59,29 @@ MIN_FULL_CYCLES_FOR_TEMPLATE = 3 MAX_CYCLES_FOR_TEMPLATE = 6 -DETECT_WINDOW_SECONDS = 15 +DETECT_WINDOW_SECONDS = 20 RECOVERY_MIN_SECONDS = 60 -HEALTHY_EMA_ALPHA = 0.12 -RECOVERY_EMA_ALPHA = 0.30 +HEALTHY_EMA_ALPHA = 0.10 +RECOVERY_EMA_ALPHA = 0.25 OUTSIDE_RATIO_THRESHOLD = 0.60 -PHASE_SEARCH_RATIO = 0.15 + VALLEY_QUANTILE = 45 -# 关键修复：预测时间轴锚定真实数据最后一个点。 -# True：预测从 last_real_ts + 1 开始，适合 Grafana 与真实曲线对齐展示。 -# False：预测从当前系统时间 + 1 开始，适合只看纯未来预测，但容易与有采集延迟的真实数据错位。 -ALIGN_PREDICTION_TO_LAST_REAL_TS = True +# phase-lock 配置 +PHASE_LOCK_MIN_WINDOW_SECONDS = 45 +PHASE_LOCK_MAX_WINDOW_SECONDS = 180 +PHASE_LOCK_PERIOD_SEARCH_RATIO = 0.12 +PHASE_LOCK_ORIGIN_SEARCH_RATIO = 0.35 +PHASE_LOCK_PERIOD_STEP = 1 +PHASE_LOCK_ORIGIN_STEP = 1 -# 如果 last_real_ts 距离当前系统时间太久，说明采集链路可能断了，跳过预测，避免用陈旧数据继续画未来线。 +# 真实数据延迟超过这个值，就不继续预测 MAX_DATA_LAG_SECONDS = 180 -# 真实数据至少推进多少秒，才写入新预测，避免同一段未来时间被反复写入。 -MIN_REAL_ADVANCE_SECONDS = 1 +# 预测锚定最后一个真实点 +ALIGN_PREDICTION_TO_LAST_REAL_TS = True # ============================================================================= @@ -130,7 +134,7 @@ ] EXTRA_PREDICT_LABELS = { - "forecast": "phase_aligned_health_v9", + "forecast": "phase_locked_health_v10", "source": "protoforge", } @@ -157,9 +161,6 @@ class BaselineState: BASELINE_STATES: Dict[str, BaselineState] = {} - -# 记录每条序列最后一次使用的真实数据时间戳，而不是预测写到哪里。 -# 这样不会把预测不断推向更远的未来。 LAST_REAL_TS_WRITTEN: Dict[str, int] = {} @@ -197,8 +198,6 @@ def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[floa return [], [] values = result[0].get("values", []) - if not values: - return [], [] ts = [] ys = [] @@ -244,6 +243,7 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. return np.array([]), np.array([]) sorted_items = sorted(data.items(), key=lambda x: x[0]) + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) ys_clean = np.array([x[1] for x in sorted_items], dtype=float) @@ -263,7 +263,7 @@ def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np. # ============================================================================= -# 周期估计与谷底检测 +# 周期估计 # ============================================================================= def moving_average(arr: np.ndarray, window: int) -> np.ndarray: @@ -355,6 +355,10 @@ def estimate_period_rough(ys_arr: np.ndarray) -> int: return int(period) +# ============================================================================= +# 谷底检测与模板构建 +# ============================================================================= + def find_valley_indices( ts_grid: np.ndarray, ys_grid: np.ndarray, @@ -366,6 +370,7 @@ def find_valley_indices( return [] period = max(3, int(expected_period)) + smooth_window = max(3, int(round(period * 0.08))) smooth_window = min(smooth_window, 21) @@ -384,6 +389,7 @@ def find_valley_indices( if len(candidates) < MIN_FULL_CYCLES_FOR_TEMPLATE: candidates = [] + for i in range(1, n - 1): if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1]: candidates.append(i) @@ -451,10 +457,6 @@ def detect_period_and_valleys( return int(period), valleys -# ============================================================================= -# 相位对齐模板 -# ============================================================================= - def build_template_from_valleys( ts_grid: np.ndarray, ys_grid: np.ndarray, @@ -546,7 +548,7 @@ def build_current_baseline( # ============================================================================= -# 预测与模板合并 +# 模板预测与重采样 # ============================================================================= def circular_template_value(template: np.ndarray, phase: float) -> float: @@ -556,6 +558,7 @@ def circular_template_value(template: np.ndarray, phase: float) -> float: return 0.0 phase = float(phase) % period + i0 = int(math.floor(phase)) % period i1 = (i0 + 1) % period frac = phase - math.floor(phase) @@ -563,46 +566,77 @@ def circular_template_value(template: np.ndarray, phase: float) -> float: return float((1.0 - frac) * template[i0] + frac * template[i1]) -def predict_with_origin( - state: BaselineState, +def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: + old_period = len(old_template) + + if old_period == new_period: + return old_template.astype(float) + + if old_period <= 1 or new_period <= 1: + return np.full(new_period, float(np.mean(old_template)), dtype=float) + + old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) + new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) + + old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) + old_y_ext = np.concatenate([old_template, old_template, old_template]) + + return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + + +def predict_template_values( + template: np.ndarray, + period: int, + phase_origin_ts: int, ts_list: List[int], - phase_origin_ts: Optional[int] = None, ) -> np.ndarray: - template = np.array(state.template, dtype=float) - period = int(state.period) - - if period <= 1 or len(template) != period: + if period <= 1: return np.zeros(len(ts_list), dtype=float) - origin = int(state.phase_origin_ts if phase_origin_ts is None else phase_origin_ts) + if len(template) != period: + template = resample_template(template, period) + values = [] for ts in ts_list: - phase = (int(ts) - origin) % period + phase = (int(ts) - int(phase_origin_ts)) % period values.append(circular_template_value(template, phase)) return np.array(values, dtype=float) -def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: - old_period = len(old_template) +def predict_with_state(state: BaselineState, ts_list: List[int]) -> np.ndarray: + template = np.array(state.template, dtype=float) - if old_period == new_period: - return old_template.astype(float) + return predict_template_values( + template=template, + period=int(state.period), + phase_origin_ts=int(state.phase_origin_ts), + ts_list=ts_list, + ) - if old_period <= 1 or new_period <= 1: - return np.full(new_period, float(np.mean(old_template)), dtype=float) - old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) - new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) +def normalize_origin_near(origin: int, period: int, near_ts: int) -> int: + if period <= 1: + return origin - old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) - old_y_ext = np.concatenate([old_template, old_template, old_template]) + origin = int(origin) + period = int(period) + near_ts = int(near_ts) - return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + while origin + period <= near_ts: + origin += period + while origin > near_ts: + origin -= period -def align_new_template_to_old(old_template: np.ndarray, new_template: np.ndarray) -> np.ndarray: + return origin + + +def align_new_template_to_old( + old_template: np.ndarray, + new_template: np.ndarray, +) -> np.ndarray: if len(old_template) != len(new_template): old_template = resample_template(old_template, len(new_template)) @@ -629,23 +663,117 @@ def align_new_template_to_old(old_template: np.ndarray, new_template: np.ndarray return best_template.astype(float) -def merge_template(old_template: np.ndarray, new_template: np.ndarray, alpha: float) -> np.ndarray: +def merge_template( + old_template: np.ndarray, + new_template: np.ndarray, + alpha: float, +) -> np.ndarray: alpha = float(np.clip(alpha, 0.0, 1.0)) if len(old_template) != len(new_template): old_template = resample_template(old_template, len(new_template)) new_template = align_new_template_to_old(old_template, new_template) + merged = (1.0 - alpha) * old_template + alpha * new_template return merged.astype(float) +# ============================================================================= +# Phase Lock +# ============================================================================= + +def phase_lock_recent( + state: BaselineState, + ts_grid: np.ndarray, + ys_grid: np.ndarray, +) -> Tuple[int, int, np.ndarray, float]: + base_period = int(state.period) + base_origin = int(state.phase_origin_ts) + base_template = np.array(state.template, dtype=float) + + if base_period <= 1 or len(base_template) <= 1: + ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() + pred = predict_with_state(state, ts_recent) + actual = ys_grid[-len(ts_recent):].astype(float) + mae = float(np.mean(np.abs(actual - pred))) if len(actual) else 0.0 + return base_period, base_origin, pred, mae + + window_seconds = max( + PHASE_LOCK_MIN_WINDOW_SECONDS, + min(PHASE_LOCK_MAX_WINDOW_SECONDS, int(base_period * 2)), + ) + + cutoff = ts_grid[-1] - window_seconds + mask = ts_grid >= cutoff + + ts_recent_arr = ts_grid[mask].astype(int) + actual = ys_grid[mask].astype(float) + + if len(ts_recent_arr) < max(10, DETECT_WINDOW_SECONDS): + ts_recent_arr = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int) + actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) + + ts_recent = ts_recent_arr.tolist() + last_ts = int(ts_recent[-1]) + + p_min = max(int(MIN_PERIOD_SECONDS), int(round(base_period * (1.0 - PHASE_LOCK_PERIOD_SEARCH_RATIO)))) + p_max = min(int(MAX_PERIOD_SECONDS), int(round(base_period * (1.0 + PHASE_LOCK_PERIOD_SEARCH_RATIO)))) + + if p_max < p_min: + p_min = p_max = base_period + + best_period = base_period + best_origin = normalize_origin_near(base_origin, base_period, last_ts) + best_template = resample_template(base_template, best_period) + best_pred = predict_template_values(best_template, best_period, best_origin, ts_recent) + best_mae = float(np.mean(np.abs(actual - best_pred))) + + for period in range(p_min, p_max + 1, PHASE_LOCK_PERIOD_STEP): + template = resample_template(base_template, period) + center_origin = normalize_origin_near(base_origin, period, last_ts) + + origin_shift = max(2, int(round(period * PHASE_LOCK_ORIGIN_SEARCH_RATIO))) + + for shift in range(-origin_shift, origin_shift + 1, PHASE_LOCK_ORIGIN_STEP): + origin = center_origin + shift + + pred = predict_template_values( + template=template, + period=period, + phase_origin_ts=origin, + ts_list=ts_recent, + ) + + mae = float(np.mean(np.abs(actual - pred))) + + # 轻微惩罚周期变化，避免过拟合抖动 + penalty = abs(period - base_period) * 0.5 + score = mae + penalty + + best_score = best_mae + abs(best_period - base_period) * 0.5 + + if score < best_score: + best_period = period + best_origin = origin + best_pred = pred + best_mae = mae + + best_origin = normalize_origin_near(best_origin, best_period, last_ts) + + return int(best_period), int(best_origin), best_pred, float(best_mae) + + # ============================================================================= # 异常检测 # ============================================================================= -def calc_threshold(pred: np.ndarray, abs_threshold: float, rel_threshold: float) -> np.ndarray: +def calc_threshold( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> np.ndarray: return np.maximum(abs_threshold, np.abs(pred) * rel_threshold) @@ -655,33 +783,8 @@ def calc_bounds( rel_threshold: float, ) -> Tuple[np.ndarray, np.ndarray]: threshold = calc_threshold(pred, abs_threshold, rel_threshold) - return pred - threshold, pred + threshold - - -def find_best_phase_origin_for_recent( - state: BaselineState, - ts_recent: List[int], - actual: np.ndarray, -) -> Tuple[int, np.ndarray, float]: - period = int(state.period) - base_origin = int(state.phase_origin_ts) - max_shift = max(1, int(round(period * PHASE_SEARCH_RATIO))) - - best_origin = base_origin - best_pred = predict_with_origin(state, ts_recent, base_origin) - best_mae = float(np.mean(np.abs(actual - best_pred))) - - for shift in range(-max_shift, max_shift + 1): - origin = base_origin + shift - pred = predict_with_origin(state, ts_recent, origin) - mae = float(np.mean(np.abs(actual - pred))) - if mae < best_mae: - best_mae = mae - best_origin = origin - best_pred = pred - - return best_origin, best_pred, best_mae + return pred - threshold, pred + threshold def detect_anomaly( @@ -690,36 +793,50 @@ def detect_anomaly( ys_grid: np.ndarray, abs_threshold: float, rel_threshold: float, -) -> Tuple[bool, float, float, float, int]: - if len(ys_grid) < DETECT_WINDOW_SECONDS: - return False, 0.0, 0.0, 0.0, int(state.phase_origin_ts) - - ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() - actual = ys_grid[-DETECT_WINDOW_SECONDS:].astype(float) - - best_origin, pred, _ = find_best_phase_origin_for_recent( +) -> Tuple[bool, float, float, float, int, int]: + best_period, best_origin, pred_recent, _ = phase_lock_recent( state=state, - ts_recent=ts_recent, - actual=actual, + ts_grid=ts_grid, + ys_grid=ys_grid, ) - threshold = calc_threshold(pred, abs_threshold, rel_threshold) - abs_err = np.abs(actual - pred) + recent_len = len(pred_recent) + + if recent_len <= 0: + return False, 0.0, 0.0, 0.0, best_period, best_origin + + actual = ys_grid[-recent_len:].astype(float) + + threshold = calc_threshold(pred_recent, abs_threshold, rel_threshold) + + abs_err = np.abs(actual - pred_recent) outside = abs_err > threshold outside_ratio = float(np.mean(outside)) mean_abs_err = float(np.mean(abs_err)) - mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1.0))) + mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred_recent), 1.0))) + is_anomaly = outside_ratio >= OUTSIDE_RATIO_THRESHOLD - return is_anomaly, outside_ratio, mean_abs_err, mean_rel_err, int(best_origin) + return ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + int(best_period), + int(best_origin), + ) # ============================================================================= # 健康基线状态管理 # ============================================================================= -def create_initial_state(ts_grid: np.ndarray, ys_grid: np.ndarray, now_sec: int) -> Optional[BaselineState]: +def create_initial_state( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + now_sec: int, +) -> Optional[BaselineState]: baseline = build_current_baseline(ts_grid, ys_grid) if baseline is None: @@ -740,6 +857,26 @@ def create_initial_state(ts_grid: np.ndarray, ys_grid: np.ndarray, now_sec: int) ) +def apply_phase_lock_to_state( + state: BaselineState, + best_period: int, + best_origin: int, +) -> None: + best_period = int(best_period) + + if best_period <= 1: + return + + template = np.array(state.template, dtype=float) + + if len(template) != best_period: + template = resample_template(template, best_period) + + state.period = best_period + state.phase_origin_ts = int(best_origin) + state.template = template.astype(float).tolist() + + def maybe_update_state( key: str, ts_grid: np.ndarray, @@ -772,7 +909,14 @@ def maybe_update_state( elapsed = min(elapsed, POLL_INTERVAL * 2) state.last_seen_ts = now_sec - is_anom, outside_ratio, mean_abs_err, mean_rel_err, best_origin = detect_anomaly( + ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + best_period, + best_origin, + ) = detect_anomaly( state=state, ts_grid=ts_grid, ys_grid=ys_grid, @@ -780,9 +924,10 @@ def maybe_update_state( rel_threshold=rel_threshold, ) - if is_anom: + if is_anomaly: state.status = BASELINE_STATUS_ANOMALY state.clean_seconds = 0 + BASELINE_STATES[key] = state logger.warning( @@ -795,13 +940,17 @@ def maybe_update_state( return state, True, outside_ratio, mean_abs_err, mean_rel_err + old_period = int(state.period) old_origin = int(state.phase_origin_ts) - state.phase_origin_ts = int(best_origin) - if abs(state.phase_origin_ts - old_origin) >= 1: - logger.debug( - "相位校正 key=%s origin %s -> %s", + apply_phase_lock_to_state(state, best_period, best_origin) + + if old_period != state.period or old_origin != state.phase_origin_ts: + logger.info( + "phase-lock key=%s period %s -> %s origin %s -> %s", key, + old_period, + state.period, datetime.fromtimestamp(old_origin).strftime("%H:%M:%S"), datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), ) @@ -809,9 +958,15 @@ def maybe_update_state( if state.status == BASELINE_STATUS_ANOMALY: state.status = BASELINE_STATUS_RECOVERING state.clean_seconds = elapsed + BASELINE_STATES[key] = state - logger.info("异常开始恢复 key=%s clean_seconds=%ss", key, state.clean_seconds) + logger.info( + "异常开始恢复 key=%s clean_seconds=%ss", + key, + state.clean_seconds, + ) + return state, False, outside_ratio, mean_abs_err, mean_rel_err if state.status == BASELINE_STATUS_RECOVERING: @@ -834,17 +989,27 @@ def maybe_update_state( int(state.period) * MAX_CYCLES_FOR_TEMPLATE, ) - baseline = build_current_baseline(ts_grid=ts_grid, ys_grid=ys_grid, tail_seconds=tail_seconds) + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_grid=ys_grid, + tail_seconds=tail_seconds, + ) if baseline is None: BASELINE_STATES[key] = state return state, False, outside_ratio, mean_abs_err, mean_rel_err new_period, new_origin, new_template = baseline + old_template = np.array(state.template, dtype=float) + alpha = RECOVERY_EMA_ALPHA if state.status == BASELINE_STATUS_RECOVERING else HEALTHY_EMA_ALPHA - merged = merge_template(old_template=old_template, new_template=new_template, alpha=alpha) + merged = merge_template( + old_template=old_template, + new_template=new_template, + alpha=alpha, + ) state.period = int(new_period) state.phase_origin_ts = int(new_origin) @@ -878,7 +1043,12 @@ def maybe_update_state( # ============================================================================= def prom_escape_label_value(value: str) -> str: - return str(value).replace("\\", "\\\\").replace("\n", "\\n").replace('"', '\\"') + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) def labels_to_str(labels: Dict[str, str]) -> str: @@ -926,7 +1096,9 @@ def write_series( resp = requests.post( f"{VM_URL}/api/v1/import/prometheus", data=payload.encode("utf-8"), - headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + headers={ + "Content-Type": "text/plain; version=0.0.4; charset=utf-8", + }, timeout=10, ) resp.raise_for_status() @@ -951,17 +1123,57 @@ def write_prediction_bundle( mean_rel_err: float, event_ts: int, ) -> bool: - ok1 = write_series(pred_metric, labels, ts_future, pred_values.astype(float).tolist()) - ok2 = write_series(f"{pred_metric}_lower", labels, ts_future, lower_values.astype(float).tolist()) - ok3 = write_series(f"{pred_metric}_upper", labels, ts_future, upper_values.astype(float).tolist()) + ok1 = write_series( + metric_name=pred_metric, + labels=labels, + ts_list=ts_future, + values=pred_values.astype(float).tolist(), + ) + + ok2 = write_series( + metric_name=f"{pred_metric}_lower", + labels=labels, + ts_list=ts_future, + values=lower_values.astype(float).tolist(), + ) + + ok3 = write_series( + metric_name=f"{pred_metric}_upper", + labels=labels, + ts_list=ts_future, + values=upper_values.astype(float).tolist(), + ) anomaly_labels = dict(labels) anomaly_labels["type"] = "prediction_deviation" - ok4 = write_series(anomaly_metric, anomaly_labels, [event_ts], [1.0 if is_anomaly else 0.0]) - ok5 = write_series(f"{anomaly_metric}_outside_ratio", anomaly_labels, [event_ts], [outside_ratio]) - ok6 = write_series(f"{anomaly_metric}_mean_abs_error", anomaly_labels, [event_ts], [mean_abs_err]) - ok7 = write_series(f"{anomaly_metric}_mean_rel_error", anomaly_labels, [event_ts], [mean_rel_err]) + ok4 = write_series( + metric_name=anomaly_metric, + labels=anomaly_labels, + ts_list=[event_ts], + values=[1.0 if is_anomaly else 0.0], + ) + + ok5 = write_series( + metric_name=f"{anomaly_metric}_outside_ratio", + labels=anomaly_labels, + ts_list=[event_ts], + values=[outside_ratio], + ) + + ok6 = write_series( + metric_name=f"{anomaly_metric}_mean_abs_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_abs_err], + ) + + ok7 = write_series( + metric_name=f"{anomaly_metric}_mean_rel_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_rel_err], + ) return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 @@ -970,7 +1182,9 @@ def write_prediction_bundle( # 标签解析 # ============================================================================= -_LABEL_PATTERN = re.compile(r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*') +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) def parse_labels_from_query(query: str) -> Dict[str, str]: @@ -987,7 +1201,14 @@ def parse_labels_from_query(query: str) -> Dict[str, str]: for match in _LABEL_PATTERN.finditer(label_part): key = match.group(1) value = match.group(2) - value = value.replace('\\"', '"').replace("\\n", "\n").replace("\\\\", "\\") + + value = ( + value + .replace('\\"', '"') + .replace("\\n", "\n") + .replace("\\\\", "\\") + ) + labels[key] = value return labels @@ -1042,7 +1263,12 @@ def load_state() -> None: states[key] = BaselineState(**value) BASELINE_STATES = states - logger.info("已加载预测状态文件 %s，状态数量=%d", STATE_FILE, len(BASELINE_STATES)) + + logger.info( + "已加载预测状态文件 %s，状态数量=%d", + STATE_FILE, + len(BASELINE_STATES), + ) except Exception as e: logger.warning("加载预测状态文件失败，将重新学习: %s", e) @@ -1069,10 +1295,14 @@ def save_state() -> None: # ============================================================================= -# 时间轴选择 +# 时间轴 # ============================================================================= -def build_prediction_timestamps(key: str, last_real_ts: int, now_sec: int) -> Optional[List[int]]: +def build_prediction_timestamps( + key: str, + last_real_ts: int, + now_sec: int, +) -> Optional[List[int]]: data_lag = now_sec - last_real_ts if data_lag > MAX_DATA_LAG_SECONDS: @@ -1086,24 +1316,24 @@ def build_prediction_timestamps(key: str, last_real_ts: int, now_sec: int) -> Op last_written_real_ts = LAST_REAL_TS_WRITTEN.get(key) - if last_written_real_ts is not None: - advance = last_real_ts - int(last_written_real_ts) - - if advance < MIN_REAL_ADVANCE_SECONDS: - logger.info( - "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written_real_ts=%s", - key, - last_real_ts, - last_written_real_ts, - ) - return None + if last_written_real_ts is not None and last_real_ts <= int(last_written_real_ts): + logger.info( + "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written_real_ts=%s", + key, + last_real_ts, + last_written_real_ts, + ) + return None if ALIGN_PREDICTION_TO_LAST_REAL_TS: base_ts = last_real_ts else: base_ts = now_sec - return [base_ts + i + 1 for i in range(WRITE_HORIZON_SECONDS)] + return [ + base_ts + i + 1 + for i in range(WRITE_HORIZON_SECONDS) + ] # ============================================================================= @@ -1123,17 +1353,28 @@ def run_once() -> None: ts, ys = fetch_history(query) if len(ys) < MIN_POINTS: - logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + logger.info( + "[%s] %s 数据不足（%d 点），跳过", + now_str, + query, + len(ys), + ) continue ts_grid, ys_grid = normalize_history(ts, ys) if len(ys_grid) < MIN_POINTS: - logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid)) + logger.info( + "[%s] %s 清洗后数据不足（%d 点），跳过", + now_str, + query, + len(ys_grid), + ) continue base_labels = parse_labels_from_query(query) write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + key = series_key(pred_metric, write_labels) state, is_anomaly, outside_ratio, mean_abs_err, mean_rel_err = maybe_update_state( @@ -1145,7 +1386,11 @@ def run_once() -> None: ) if state is None: - logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + logger.info( + "[%s] %s 暂无可用健康模板，等待学习", + now_str, + query, + ) continue now_sec = int(time.time()) @@ -1161,7 +1406,8 @@ def run_once() -> None: if not ts_future: continue - pred_values = predict_with_origin(state, ts_future) + pred_values = predict_with_state(state, ts_future) + lower_values, upper_values = calc_bounds( pred=pred_values, abs_threshold=abs_threshold, @@ -1184,7 +1430,11 @@ def run_once() -> None: ) if not ok: - logger.error("[%s] %s 写入预测数据失败", now_str, query) + logger.error( + "[%s] %s 写入预测数据失败", + now_str, + query, + ) continue LAST_REAL_TS_WRITTEN[key] = last_real_ts @@ -1234,4 +1484,4 @@ def main() -> None: if __name__ == "__main__": - main() + main() \ No newline at end of file From 79e9f9b080e3f5fc4a284ab1422d3a54aacaff0e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 21:23:40 +0800 Subject: [PATCH 26/36] feat(pridict_v4): update pridict v4 version --- ai/pridict_v4.py | 1604 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 1604 insertions(+) create mode 100644 ai/pridict_v4.py diff --git a/ai/pridict_v4.py b/ai/pridict_v4.py new file mode 100644 index 0000000..774ad3a --- /dev/null +++ b/ai/pridict_v4.py @@ -0,0 +1,1604 @@ +# -*- coding: utf-8 -*- +""" +ProtoForge Predictor v11 + +核心能力： +1. feed_rate / spindle_speed / spindle_current 使用 phase-lock 点预测。 +2. vibration_x / vibration_y / vibration_z 使用 phase-band 预测带。 +3. vibration 类指标不再追求单点完全贴合，而是输出： + - xxx_predicted 中位数预测线 + - xxx_predicted_upper 正常上边界 + - xxx_predicted_lower 正常下边界 +4. 预测起点锚定最后一个真实点 last_real_ts，避免时间错位。 +5. 异常期间冻结健康模板，不学习故障数据。 +6. 故障恢复后等待稳定，再恢复模板学习。 +""" + +import json +import logging +import math +import os +import re +import time +from dataclasses import asdict, dataclass +from datetime import datetime, timedelta +from typing import Dict, List, Optional, Tuple + +import numpy as np +import requests + + +# ============================================================================= +# 日志配置 +# ============================================================================= + +logging.basicConfig( + level=logging.INFO, + format="%(asctime)s [%(levelname)s] %(message)s", +) + +logger = logging.getLogger(__name__) + + +# ============================================================================= +# 基础配置 +# ============================================================================= + +VM_URL = "http://localhost:8428" +STATE_FILE = "/tmp/protoforge_predictor_state_v11.json" + +HISTORY_MINUTES = 30 +HORIZON_SECONDS = 120 +POLL_INTERVAL = 30 + +WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) + +QUERY_STEP = "1s" +MIN_POINTS = 120 + +MIN_PERIOD_SECONDS = 5 +MAX_PERIOD_SECONDS = 3600 + +MIN_FULL_CYCLES_FOR_TEMPLATE = 3 +MAX_CYCLES_FOR_TEMPLATE = 8 + +DETECT_WINDOW_SECONDS = 20 +RECOVERY_MIN_SECONDS = 60 + +HEALTHY_EMA_ALPHA = 0.10 +RECOVERY_EMA_ALPHA = 0.25 + +OUTSIDE_RATIO_THRESHOLD = 0.60 + +VALLEY_QUANTILE = 45 + +MAX_DATA_LAG_SECONDS = 180 + +PHASE_LOCK_MIN_WINDOW_SECONDS = 45 +PHASE_LOCK_MAX_WINDOW_SECONDS = 180 +PHASE_LOCK_PERIOD_SEARCH_RATIO = 0.12 +PHASE_LOCK_ORIGIN_SEARCH_RATIO = 0.35 +PHASE_LOCK_PERIOD_STEP = 1 +PHASE_LOCK_ORIGIN_STEP = 1 + + +# ============================================================================= +# 指标配置 +# ============================================================================= + +PREDICT_TARGETS = [ + { + "query": 'feed_rate{device_id="fanuc-cnc"}', + "pred_metric": "feed_rate_predicted", + "anomaly_metric": "feed_rate_anomaly", + "strategy": "phase_point", + "abs_threshold": 400.0, + "rel_threshold": 0.25, + "smooth_window": 1, + }, + { + "query": 'spindle_speed{device_id="fanuc-cnc"}', + "pred_metric": "spindle_speed_predicted", + "anomaly_metric": "spindle_speed_anomaly", + "strategy": "phase_point", + "abs_threshold": 500.0, + "rel_threshold": 0.25, + "smooth_window": 1, + }, + { + "query": 'spindle_current{device_id="fanuc-cnc"}', + "pred_metric": "spindle_current_predicted", + "anomaly_metric": "spindle_current_anomaly", + "strategy": "phase_point", + "abs_threshold": 5.0, + "rel_threshold": 0.25, + "smooth_window": 1, + }, + { + "query": 'vibration_x{device_id="fanuc-cnc"}', + "pred_metric": "vibration_x_predicted", + "anomaly_metric": "vibration_x_anomaly", + "strategy": "phase_band", + "abs_threshold": 0.12, + "rel_threshold": 0.35, + "smooth_window": 5, + "band_low_q": 10, + "band_high_q": 90, + "band_pad_abs": 0.06, + }, + { + "query": 'vibration_y{device_id="fanuc-cnc"}', + "pred_metric": "vibration_y_predicted", + "anomaly_metric": "vibration_y_anomaly", + "strategy": "phase_band", + "abs_threshold": 0.12, + "rel_threshold": 0.35, + "smooth_window": 5, + "band_low_q": 10, + "band_high_q": 90, + "band_pad_abs": 0.06, + }, + { + "query": 'vibration_z{device_id="fanuc-cnc"}', + "pred_metric": "vibration_z_predicted", + "anomaly_metric": "vibration_z_anomaly", + "strategy": "phase_band", + "abs_threshold": 0.12, + "rel_threshold": 0.35, + "smooth_window": 5, + "band_low_q": 10, + "band_high_q": 90, + "band_pad_abs": 0.06, + }, +] + +EXTRA_PREDICT_LABELS = { + "forecast": "phase_band_health_v11", + "source": "protoforge", +} + +BASELINE_STATUS_HEALTHY = "healthy" +BASELINE_STATUS_ANOMALY = "anomaly" +BASELINE_STATUS_RECOVERING = "recovering" + + +# ============================================================================= +# 状态结构 +# ============================================================================= + +@dataclass +class BaselineState: + period: int + phase_origin_ts: int + template: List[float] + lower_template: List[float] + upper_template: List[float] + strategy: str + status: str + clean_seconds: int + last_update_ts: int + last_seen_ts: int + y_min: float + y_max: float + + +BASELINE_STATES: Dict[str, BaselineState] = {} +LAST_REAL_TS_WRITTEN: Dict[str, int] = {} + + +# ============================================================================= +# VictoriaMetrics 读取 +# ============================================================================= + +def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: + now = datetime.now() + start = now - timedelta(minutes=minutes) + + try: + resp = requests.get( + f"{VM_URL}/api/v1/query_range", + params={ + "query": query, + "start": start.timestamp(), + "end": now.timestamp(), + "step": QUERY_STEP, + }, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("拉取数据失败 query=%s: %s", query, e) + return [], [] + + try: + result = resp.json().get("data", {}).get("result", []) + except Exception as e: + logger.error("解析 VM 返回失败 query=%s: %s", query, e) + return [], [] + + if not result: + return [], [] + + values = result[0].get("values", []) + + ts = [] + ys = [] + + for item in values: + if len(item) < 2: + continue + + try: + t = float(item[0]) + y = float(item[1]) + except Exception: + continue + + if not math.isfinite(t) or not math.isfinite(y): + continue + + ts.append(t) + ys.append(y) + + return ts, ys + + +def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: + if not ts or not ys or len(ts) != len(ys): + return np.array([]), np.array([]) + + data = {} + + for t, y in zip(ts, ys): + try: + sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(sec) or not math.isfinite(val): + continue + + data[sec] = val + + if not data: + return np.array([]), np.array([]) + + sorted_items = sorted(data.items(), key=lambda x: x[0]) + + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) + ys_clean = np.array([x[1] for x in sorted_items], dtype=float) + + if len(ts_clean) < 2: + return ts_clean, ys_clean + + start_sec = int(ts_clean[0]) + end_sec = int(ts_clean[-1]) + + if end_sec <= start_sec: + return ts_clean, ys_clean + + ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) + ys_grid = np.interp(ts_grid, ts_clean, ys_clean) + + return ts_grid, ys_grid + + +# ============================================================================= +# 平滑与预处理 +# ============================================================================= + +def rolling_median(arr: np.ndarray, window: int) -> np.ndarray: + if window <= 1 or len(arr) < window: + return arr.astype(float) + + if window % 2 == 0: + window += 1 + + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + result = [] + + for i in range(len(arr)): + result.append(float(np.median(padded[i:i + window]))) + + return np.array(result, dtype=float) + + +def moving_average(arr: np.ndarray, window: int) -> np.ndarray: + if window <= 1 or len(arr) < window: + return arr.astype(float) + + if window % 2 == 0: + window += 1 + + kernel = np.ones(window, dtype=float) / window + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + return np.convolve(padded, kernel, mode="valid") + + +def preprocess_values(ys_grid: np.ndarray, target: Dict) -> np.ndarray: + strategy = target.get("strategy", "phase_point") + smooth_window = int(target.get("smooth_window", 1)) + + if strategy == "phase_band": + return rolling_median(ys_grid, smooth_window) + + if smooth_window > 1: + return moving_average(ys_grid, smooth_window) + + return ys_grid.astype(float) + + +# ============================================================================= +# 周期估计 +# ============================================================================= + +def estimate_period_by_fft(ys_arr: np.ndarray) -> float: + n = len(ys_arr) + + if n < 8: + return 60.0 + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return 60.0 + + fft_vals = np.fft.rfft(centered) + freqs = np.fft.rfftfreq(n, d=1.0) + + if len(freqs) <= 1: + return 60.0 + + power = np.abs(fft_vals[1:]) + + if len(power) == 0 or np.max(power) <= 0: + return 60.0 + + dominant_idx = int(np.argmax(power)) + 1 + dominant_freq = float(freqs[dominant_idx]) + + if dominant_freq <= 0: + return 60.0 + + period = 1.0 / dominant_freq + + return float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: + n = len(ys_arr) + + if n < 20: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + corr = np.correlate(centered, centered, mode="full")[n - 1:] + + p0 = int(round(init_period)) + left = max(int(MIN_PERIOD_SECONDS), int(max(2, p0 * 0.7))) + right = min(n // 2, int(max(left + 1, p0 * 1.3))) + + if right <= left: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + search = corr[left:right + 1] + + if len(search) == 0: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + best_lag = left + int(np.argmax(search)) + + return float(np.clip(best_lag, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def estimate_period_rough(ys_arr: np.ndarray) -> int: + p_fft = estimate_period_by_fft(ys_arr) + p_refined = refine_period_by_autocorr(ys_arr, p_fft) + + period = int(round(p_refined)) + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) + + return int(period) + + +# ============================================================================= +# 谷底检测 +# ============================================================================= + +def find_valley_indices( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + expected_period: int, +) -> List[int]: + n = len(ys_grid) + + if n < max(10, expected_period * 2): + return [] + + period = max(3, int(expected_period)) + smooth_window = max(3, int(round(period * 0.08))) + smooth_window = min(smooth_window, 21) + + ys_smooth = moving_average(ys_grid, smooth_window) + threshold = float(np.percentile(ys_smooth, VALLEY_QUANTILE)) + + candidates = [] + + for i in range(1, n - 1): + if ( + ys_smooth[i] <= ys_smooth[i - 1] + and ys_smooth[i] < ys_smooth[i + 1] + and ys_smooth[i] <= threshold + ): + candidates.append(i) + + if len(candidates) < MIN_FULL_CYCLES_FOR_TEMPLATE: + candidates = [] + + for i in range(1, n - 1): + if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1]: + candidates.append(i) + + if not candidates: + return [] + + min_distance = max(2, int(round(period * 0.55))) + selected = [] + + for idx in candidates: + if not selected: + selected.append(idx) + continue + + if idx - selected[-1] >= min_distance: + selected.append(idx) + continue + + if ys_smooth[idx] < ys_smooth[selected[-1]]: + selected[-1] = idx + + if len(selected) < 2: + return selected + + cleaned = [selected[0]] + + for idx in selected[1:]: + diff = int(ts_grid[idx] - ts_grid[cleaned[-1]]) + + if int(period * 0.55) <= diff <= int(period * 1.60): + cleaned.append(idx) + continue + + if diff < int(period * 0.55): + if ys_smooth[idx] < ys_smooth[cleaned[-1]]: + cleaned[-1] = idx + continue + + cleaned.append(idx) + + return cleaned + + +def detect_period_and_valleys( + ts_grid: np.ndarray, + ys_grid: np.ndarray, +) -> Tuple[int, List[int]]: + rough = estimate_period_rough(ys_grid) + valleys = find_valley_indices(ts_grid, ys_grid, rough) + + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys]) + good = diffs[(diffs >= rough * 0.55) & (diffs <= rough * 1.60)] + + if len(good) > 0: + period = int(round(float(np.median(good)))) + else: + period = rough + else: + period = rough + + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) + + return int(period), valleys + + +# ============================================================================= +# 模板构建 +# ============================================================================= + +def build_templates_from_valleys( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + period: int, + valleys: List[int], + target: Dict, +) -> Optional[Tuple[np.ndarray, np.ndarray, np.ndarray]]: + if period <= 1 or len(valleys) < MIN_FULL_CYCLES_FOR_TEMPLATE + 1: + return None + + strategy = target.get("strategy", "phase_point") + low_q = float(target.get("band_low_q", 10)) + high_q = float(target.get("band_high_q", 90)) + + pairs = [] + + for a, b in zip(valleys[:-1], valleys[1:]): + cycle_len = float(ts_grid[b] - ts_grid[a]) + + if period * 0.55 <= cycle_len <= period * 1.60: + pairs.append((a, b, cycle_len)) + + if len(pairs) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + pairs = pairs[-MAX_CYCLES_FOR_TEMPLATE:] + + phase_grid = np.arange(period, dtype=float) + segments = [] + weights = [] + + for idx, (a, b, cycle_len) in enumerate(pairs): + seg_ts = ts_grid[a:b + 1] + seg_y = ys_grid[a:b + 1] + + if len(seg_y) < 3: + continue + + x_old = (seg_ts - seg_ts[0]) / cycle_len * period + seg = np.interp(phase_grid, x_old, seg_y) + + segments.append(seg.astype(float)) + weights.append(0.5 + 0.5 * ((idx + 1) / len(pairs))) + + if len(segments) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + arr = np.vstack(segments) + w_arr = np.array(weights, dtype=float) + + if strategy == "phase_band": + mid_template = np.percentile(arr, 50, axis=0) + lower_template = np.percentile(arr, low_q, axis=0) + upper_template = np.percentile(arr, high_q, axis=0) + else: + mid_template = np.average(arr, axis=0, weights=w_arr) + lower_template = mid_template.copy() + upper_template = mid_template.copy() + + return ( + mid_template.astype(float), + lower_template.astype(float), + upper_template.astype(float), + ) + + +def build_current_baseline( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + target: Dict, + tail_seconds: Optional[int] = None, +) -> Optional[Tuple[int, int, np.ndarray, np.ndarray, np.ndarray]]: + if len(ys_grid) < MIN_POINTS: + return None + + if tail_seconds is not None and tail_seconds > 0: + cutoff = ts_grid[-1] - int(tail_seconds) + mask = ts_grid >= cutoff + ts_use = ts_grid[mask] + ys_use = ys_grid[mask] + else: + ts_use = ts_grid + ys_use = ys_grid + + if len(ys_use) < MIN_POINTS: + return None + + period, valleys = detect_period_and_valleys(ts_use, ys_use) + + templates = build_templates_from_valleys( + ts_grid=ts_use, + ys_grid=ys_use, + period=period, + valleys=valleys, + target=target, + ) + + if templates is None or len(valleys) == 0: + return None + + template, lower_template, upper_template = templates + phase_origin_ts = int(round(float(ts_use[valleys[-1]]))) + + return int(period), phase_origin_ts, template, lower_template, upper_template + + +# ============================================================================= +# 模板预测 +# ============================================================================= + +def circular_template_value(template: np.ndarray, phase: float) -> float: + period = len(template) + + if period == 0: + return 0.0 + + phase = float(phase) % period + i0 = int(math.floor(phase)) % period + i1 = (i0 + 1) % period + frac = phase - math.floor(phase) + + return float((1.0 - frac) * template[i0] + frac * template[i1]) + + +def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: + old_period = len(old_template) + + if old_period == new_period: + return old_template.astype(float) + + if old_period <= 1 or new_period <= 1: + return np.full(new_period, float(np.mean(old_template)), dtype=float) + + old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) + new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) + + old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) + old_y_ext = np.concatenate([old_template, old_template, old_template]) + + return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + + +def predict_template_values( + template: np.ndarray, + period: int, + phase_origin_ts: int, + ts_list: List[int], +) -> np.ndarray: + if period <= 1: + return np.zeros(len(ts_list), dtype=float) + + if len(template) != period: + template = resample_template(template, period) + + values = [] + + for ts in ts_list: + phase = (int(ts) - int(phase_origin_ts)) % period + values.append(circular_template_value(template, phase)) + + return np.array(values, dtype=float) + + +def predict_state_bundle( + state: BaselineState, + ts_list: List[int], +) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: + period = int(state.period) + origin = int(state.phase_origin_ts) + + mid = predict_template_values( + template=np.array(state.template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + lower = predict_template_values( + template=np.array(state.lower_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + upper = predict_template_values( + template=np.array(state.upper_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + return mid, lower, upper + + +def normalize_origin_near(origin: int, period: int, near_ts: int) -> int: + if period <= 1: + return origin + + origin = int(origin) + period = int(period) + near_ts = int(near_ts) + + while origin + period <= near_ts: + origin += period + + while origin > near_ts: + origin -= period + + return origin + + +def merge_template( + old_template: np.ndarray, + new_template: np.ndarray, + alpha: float, +) -> np.ndarray: + alpha = float(np.clip(alpha, 0.0, 1.0)) + + if len(old_template) != len(new_template): + old_template = resample_template(old_template, len(new_template)) + + merged = (1.0 - alpha) * old_template + alpha * new_template + + return merged.astype(float) + + +# ============================================================================= +# Phase Lock +# ============================================================================= + +def phase_lock_recent( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, +) -> Tuple[int, int, np.ndarray, float]: + base_period = int(state.period) + base_origin = int(state.phase_origin_ts) + base_template = np.array(state.template, dtype=float) + + if base_period <= 1 or len(base_template) <= 1: + ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() + pred = predict_template_values(base_template, base_period, base_origin, ts_recent) + actual = ys_model[-len(ts_recent):].astype(float) + mae = float(np.mean(np.abs(actual - pred))) if len(actual) else 0.0 + return base_period, base_origin, pred, mae + + window_seconds = max( + PHASE_LOCK_MIN_WINDOW_SECONDS, + min(PHASE_LOCK_MAX_WINDOW_SECONDS, int(base_period * 2)), + ) + + cutoff = ts_grid[-1] - window_seconds + mask = ts_grid >= cutoff + + ts_recent_arr = ts_grid[mask].astype(int) + actual = ys_model[mask].astype(float) + + if len(ts_recent_arr) < max(10, DETECT_WINDOW_SECONDS): + ts_recent_arr = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int) + actual = ys_model[-DETECT_WINDOW_SECONDS:].astype(float) + + ts_recent = ts_recent_arr.tolist() + last_ts = int(ts_recent[-1]) + + p_min = max( + int(MIN_PERIOD_SECONDS), + int(round(base_period * (1.0 - PHASE_LOCK_PERIOD_SEARCH_RATIO))), + ) + p_max = min( + int(MAX_PERIOD_SECONDS), + int(round(base_period * (1.0 + PHASE_LOCK_PERIOD_SEARCH_RATIO))), + ) + + best_period = base_period + best_origin = normalize_origin_near(base_origin, base_period, last_ts) + best_template = resample_template(base_template, best_period) + + best_pred = predict_template_values( + template=best_template, + period=best_period, + phase_origin_ts=best_origin, + ts_list=ts_recent, + ) + + best_mae = float(np.mean(np.abs(actual - best_pred))) + + for period in range(p_min, p_max + 1, PHASE_LOCK_PERIOD_STEP): + template = resample_template(base_template, period) + center_origin = normalize_origin_near(base_origin, period, last_ts) + origin_shift = max(2, int(round(period * PHASE_LOCK_ORIGIN_SEARCH_RATIO))) + + for shift in range(-origin_shift, origin_shift + 1, PHASE_LOCK_ORIGIN_STEP): + origin = center_origin + shift + + pred = predict_template_values( + template=template, + period=period, + phase_origin_ts=origin, + ts_list=ts_recent, + ) + + mae = float(np.mean(np.abs(actual - pred))) + penalty = abs(period - base_period) * 0.5 + score = mae + penalty + + best_score = best_mae + abs(best_period - base_period) * 0.5 + + if score < best_score: + best_period = period + best_origin = origin + best_pred = pred + best_mae = mae + + best_origin = normalize_origin_near(best_origin, best_period, last_ts) + + return int(best_period), int(best_origin), best_pred, float(best_mae) + + +# ============================================================================= +# 异常检测 +# ============================================================================= + +def calc_point_bounds( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[np.ndarray, np.ndarray]: + threshold = np.maximum(abs_threshold, np.abs(pred) * rel_threshold) + return pred - threshold, pred + threshold + + +def calc_final_bounds( + state: BaselineState, + pred: np.ndarray, + lower_raw: np.ndarray, + upper_raw: np.ndarray, + target: Dict, +) -> Tuple[np.ndarray, np.ndarray]: + strategy = target.get("strategy", "phase_point") + abs_threshold = float(target.get("abs_threshold", 1.0)) + rel_threshold = float(target.get("rel_threshold", 0.25)) + + if strategy == "phase_band": + pad_abs = float(target.get("band_pad_abs", abs_threshold)) + dynamic_pad = np.maximum(pad_abs, np.abs(pred) * rel_threshold * 0.20) + lower = lower_raw - dynamic_pad + upper = upper_raw + dynamic_pad + return lower, upper + + return calc_point_bounds(pred, abs_threshold, rel_threshold) + + +def detect_anomaly( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, + target: Dict, +) -> Tuple[bool, float, float, float, int, int]: + best_period, best_origin, pred_recent, _ = phase_lock_recent( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + ) + + recent_len = len(pred_recent) + + if recent_len <= 0: + return False, 0.0, 0.0, 0.0, best_period, best_origin + + actual = ys_model[-recent_len:].astype(float) + + tmp_state = BaselineState( + period=best_period, + phase_origin_ts=best_origin, + template=state.template, + lower_template=state.lower_template, + upper_template=state.upper_template, + strategy=state.strategy, + status=state.status, + clean_seconds=state.clean_seconds, + last_update_ts=state.last_update_ts, + last_seen_ts=state.last_seen_ts, + y_min=state.y_min, + y_max=state.y_max, + ) + + recent_ts = ts_grid[-recent_len:].astype(int).tolist() + pred, lower_raw, upper_raw = predict_state_bundle(tmp_state, recent_ts) + + lower, upper = calc_final_bounds( + state=tmp_state, + pred=pred, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + outside = (actual < lower) | (actual > upper) + abs_err = np.abs(actual - pred) + + outside_ratio = float(np.mean(outside)) + mean_abs_err = float(np.mean(abs_err)) + mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1e-6))) + + is_anomaly = outside_ratio >= OUTSIDE_RATIO_THRESHOLD + + return ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + int(best_period), + int(best_origin), + ) + + +# ============================================================================= +# 状态管理 +# ============================================================================= + +def create_initial_state( + ts_grid: np.ndarray, + ys_model: np.ndarray, + target: Dict, + now_sec: int, +) -> Optional[BaselineState]: + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_grid=ys_model, + target=target, + ) + + if baseline is None: + return None + + period, phase_origin_ts, template, lower_template, upper_template = baseline + + return BaselineState( + period=int(period), + phase_origin_ts=int(phase_origin_ts), + template=template.astype(float).tolist(), + lower_template=lower_template.astype(float).tolist(), + upper_template=upper_template.astype(float).tolist(), + strategy=str(target.get("strategy", "phase_point")), + status=BASELINE_STATUS_HEALTHY, + clean_seconds=int(period * MAX_CYCLES_FOR_TEMPLATE), + last_update_ts=now_sec, + last_seen_ts=now_sec, + y_min=float(np.min(ys_model)), + y_max=float(np.max(ys_model)), + ) + + +def apply_phase_lock_to_state( + state: BaselineState, + best_period: int, + best_origin: int, +) -> None: + best_period = int(best_period) + + if best_period <= 1: + return + + if len(state.template) != best_period: + state.template = resample_template( + np.array(state.template, dtype=float), + best_period, + ).astype(float).tolist() + + if len(state.lower_template) != best_period: + state.lower_template = resample_template( + np.array(state.lower_template, dtype=float), + best_period, + ).astype(float).tolist() + + if len(state.upper_template) != best_period: + state.upper_template = resample_template( + np.array(state.upper_template, dtype=float), + best_period, + ).astype(float).tolist() + + state.period = best_period + state.phase_origin_ts = int(best_origin) + + +def maybe_update_state( + key: str, + ts_grid: np.ndarray, + ys_model: np.ndarray, + target: Dict, +) -> Tuple[Optional[BaselineState], bool, float, float, float]: + now_sec = int(time.time()) + state = BASELINE_STATES.get(key) + + if state is None: + state = create_initial_state( + ts_grid=ts_grid, + ys_model=ys_model, + target=target, + now_sec=now_sec, + ) + + if state is None: + return None, False, 0.0, 0.0, 0.0 + + BASELINE_STATES[key] = state + + logger.info( + "初始化健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss", + key, + state.strategy, + state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, + ) + + return state, False, 0.0, 0.0, 0.0 + + elapsed = max(1, now_sec - int(state.last_seen_ts)) + elapsed = min(elapsed, POLL_INTERVAL * 2) + state.last_seen_ts = now_sec + + ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + best_period, + best_origin, + ) = detect_anomaly( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + target=target, + ) + + if is_anomaly: + state.status = BASELINE_STATUS_ANOMALY + state.clean_seconds = 0 + BASELINE_STATES[key] = state + + logger.warning( + "检测到异常，冻结模板 key=%s outside_ratio=%.2f mean_abs_err=%.4f mean_rel_err=%.4f", + key, + outside_ratio, + mean_abs_err, + mean_rel_err, + ) + + return state, True, outside_ratio, mean_abs_err, mean_rel_err + + old_period = int(state.period) + old_origin = int(state.phase_origin_ts) + + apply_phase_lock_to_state(state, best_period, best_origin) + + if old_period != state.period or old_origin != state.phase_origin_ts: + logger.info( + "phase-lock key=%s period %s -> %s origin %s -> %s", + key, + old_period, + state.period, + datetime.fromtimestamp(old_origin).strftime("%H:%M:%S"), + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + ) + + if state.status == BASELINE_STATUS_ANOMALY: + state.status = BASELINE_STATUS_RECOVERING + state.clean_seconds = elapsed + BASELINE_STATES[key] = state + + logger.info( + "异常开始恢复 key=%s clean_seconds=%ss", + key, + state.clean_seconds, + ) + + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + if state.status == BASELINE_STATUS_RECOVERING: + state.clean_seconds += elapsed + else: + state.status = BASELINE_STATUS_HEALTHY + state.clean_seconds += elapsed + + min_clean_for_update = max( + RECOVERY_MIN_SECONDS, + int(state.period) * MIN_FULL_CYCLES_FOR_TEMPLATE, + ) + + if state.clean_seconds < min_clean_for_update: + BASELINE_STATES[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + tail_seconds = min( + int(state.clean_seconds), + int(state.period) * MAX_CYCLES_FOR_TEMPLATE, + ) + + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_grid=ys_model, + target=target, + tail_seconds=tail_seconds, + ) + + if baseline is None: + BASELINE_STATES[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + new_period, new_origin, new_template, new_lower_template, new_upper_template = baseline + + alpha = RECOVERY_EMA_ALPHA if state.status == BASELINE_STATUS_RECOVERING else HEALTHY_EMA_ALPHA + + state.template = merge_template( + np.array(state.template, dtype=float), + new_template, + alpha, + ).astype(float).tolist() + + state.lower_template = merge_template( + np.array(state.lower_template, dtype=float), + new_lower_template, + alpha, + ).astype(float).tolist() + + state.upper_template = merge_template( + np.array(state.upper_template, dtype=float), + new_upper_template, + alpha, + ).astype(float).tolist() + + state.period = int(new_period) + state.phase_origin_ts = int(new_origin) + state.status = BASELINE_STATUS_HEALTHY + state.last_update_ts = now_sec + + if tail_seconds > 0 and len(ys_model) >= tail_seconds: + state.y_min = float(np.min(ys_model[-tail_seconds:])) + state.y_max = float(np.max(ys_model[-tail_seconds:])) + else: + state.y_min = float(np.min(ys_model)) + state.y_max = float(np.max(ys_model)) + + BASELINE_STATES[key] = state + + logger.info( + "更新健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss alpha=%.2f", + key, + state.strategy, + state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, + alpha, + ) + + return state, False, outside_ratio, mean_abs_err, mean_rel_err + + +# ============================================================================= +# Prometheus 写入 +# ============================================================================= + +def prom_escape_label_value(value: str) -> str: + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) + + +def labels_to_str(labels: Dict[str, str]) -> str: + if not labels: + return "" + + parts = [] + + for k in sorted(labels.keys()): + parts.append(f'{k}="{prom_escape_label_value(labels[k])}"') + + return "{" + ",".join(parts) + "}" + + +def write_series( + metric_name: str, + labels: Dict[str, str], + ts_list: List[int], + values: List[float], +) -> bool: + if not ts_list or not values or len(ts_list) != len(values): + return False + + label_str = labels_to_str(labels) + lines = [] + + for t, y in zip(ts_list, values): + try: + ts_sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(ts_sec) or not math.isfinite(val): + continue + + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_sec * 1000}") + + if not lines: + return False + + payload = "\n".join(lines) + "\n" + + try: + resp = requests.post( + f"{VM_URL}/api/v1/import/prometheus", + data=payload.encode("utf-8"), + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + timeout=10, + ) + resp.raise_for_status() + return True + + except requests.RequestException as e: + logger.error("写入数据失败 metric=%s: %s", metric_name, e) + return False + + +def write_prediction_bundle( + pred_metric: str, + anomaly_metric: str, + labels: Dict[str, str], + ts_future: List[int], + pred_values: np.ndarray, + lower_values: np.ndarray, + upper_values: np.ndarray, + is_anomaly: bool, + outside_ratio: float, + mean_abs_err: float, + mean_rel_err: float, + event_ts: int, +) -> bool: + ok1 = write_series( + metric_name=pred_metric, + labels=labels, + ts_list=ts_future, + values=pred_values.astype(float).tolist(), + ) + + ok2 = write_series( + metric_name=f"{pred_metric}_lower", + labels=labels, + ts_list=ts_future, + values=lower_values.astype(float).tolist(), + ) + + ok3 = write_series( + metric_name=f"{pred_metric}_upper", + labels=labels, + ts_list=ts_future, + values=upper_values.astype(float).tolist(), + ) + + anomaly_labels = dict(labels) + anomaly_labels["type"] = "prediction_deviation" + + ok4 = write_series( + metric_name=anomaly_metric, + labels=anomaly_labels, + ts_list=[event_ts], + values=[1.0 if is_anomaly else 0.0], + ) + + ok5 = write_series( + metric_name=f"{anomaly_metric}_outside_ratio", + labels=anomaly_labels, + ts_list=[event_ts], + values=[outside_ratio], + ) + + ok6 = write_series( + metric_name=f"{anomaly_metric}_mean_abs_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_abs_err], + ) + + ok7 = write_series( + metric_name=f"{anomaly_metric}_mean_rel_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_rel_err], + ) + + return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 + + +# ============================================================================= +# 标签解析 +# ============================================================================= + +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) + + +def parse_labels_from_query(query: str) -> Dict[str, str]: + labels = {} + + if "{" not in query or "}" not in query: + return labels + + try: + label_part = query[query.index("{") + 1:query.rindex("}")] + except Exception: + return labels + + for match in _LABEL_PATTERN.finditer(label_part): + key = match.group(1) + value = match.group(2) + + value = ( + value + .replace('\\"', '"') + .replace("\\n", "\n") + .replace("\\\\", "\\") + ) + + labels[key] = value + + return labels + + +def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: + result = {} + + for d in dicts: + if d: + result.update(d) + + return result + + +def series_key(metric_name: str, labels: Dict[str, str]) -> str: + return metric_name + labels_to_str(labels) + + +# ============================================================================= +# 状态持久化 +# ============================================================================= + +def load_state() -> None: + global BASELINE_STATES + + if not os.path.exists(STATE_FILE): + return + + try: + with open(STATE_FILE, "r", encoding="utf-8") as f: + raw = json.load(f) + + states = {} + + for key, value in raw.get("baseline_states", {}).items(): + required_fields = { + "period", + "phase_origin_ts", + "template", + "lower_template", + "upper_template", + "strategy", + "status", + "clean_seconds", + "last_update_ts", + "last_seen_ts", + "y_min", + "y_max", + } + + if not required_fields.issubset(set(value.keys())): + continue + + states[key] = BaselineState(**value) + + BASELINE_STATES = states + + logger.info( + "已加载预测状态文件 %s，状态数量=%d", + STATE_FILE, + len(BASELINE_STATES), + ) + + except Exception as e: + logger.warning("加载预测状态文件失败，将重新学习: %s", e) + + +def save_state() -> None: + try: + raw = { + "baseline_states": { + key: asdict(value) + for key, value in BASELINE_STATES.items() + } + } + + tmp_file = STATE_FILE + ".tmp" + + with open(tmp_file, "w", encoding="utf-8") as f: + json.dump(raw, f, ensure_ascii=False, indent=2) + + os.replace(tmp_file, STATE_FILE) + + except Exception as e: + logger.warning("保存预测状态文件失败: %s", e) + + +# ============================================================================= +# 时间轴 +# ============================================================================= + +def build_prediction_timestamps( + key: str, + last_real_ts: int, + now_sec: int, +) -> Optional[List[int]]: + data_lag = now_sec - last_real_ts + + if data_lag > MAX_DATA_LAG_SECONDS: + logger.warning( + "真实数据延迟过大，跳过预测 key=%s data_lag=%ss max=%ss", + key, + data_lag, + MAX_DATA_LAG_SECONDS, + ) + return None + + last_written_real_ts = LAST_REAL_TS_WRITTEN.get(key) + + if last_written_real_ts is not None and last_real_ts <= int(last_written_real_ts): + logger.info( + "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written_real_ts=%s", + key, + last_real_ts, + last_written_real_ts, + ) + return None + + base_ts = last_real_ts + + return [ + base_ts + i + 1 + for i in range(WRITE_HORIZON_SECONDS) + ] + + +# ============================================================================= +# 主流程 +# ============================================================================= + +def run_once() -> None: + now_str = datetime.now().strftime("%H:%M:%S") + + for target in PREDICT_TARGETS: + query = target["query"] + pred_metric = target["pred_metric"] + anomaly_metric = target["anomaly_metric"] + + ts, ys = fetch_history(query) + + if len(ys) < MIN_POINTS: + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + continue + + ts_grid, ys_grid_raw = normalize_history(ts, ys) + + if len(ys_grid_raw) < MIN_POINTS: + logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid_raw)) + continue + + ys_grid_model = preprocess_values(ys_grid_raw, target) + + base_labels = parse_labels_from_query(query) + write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + + key = series_key(pred_metric, write_labels) + + state, is_anomaly, outside_ratio, mean_abs_err, mean_rel_err = maybe_update_state( + key=key, + ts_grid=ts_grid, + ys_model=ys_grid_model, + target=target, + ) + + if state is None: + logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + continue + + now_sec = int(time.time()) + last_real_ts = int(ts_grid[-1]) + data_lag = now_sec - last_real_ts + + ts_future = build_prediction_timestamps( + key=key, + last_real_ts=last_real_ts, + now_sec=now_sec, + ) + + if not ts_future: + continue + + pred_values, lower_raw, upper_raw = predict_state_bundle(state, ts_future) + + lower_values, upper_values = calc_final_bounds( + state=state, + pred=pred_values, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + ok = write_prediction_bundle( + pred_metric=pred_metric, + anomaly_metric=anomaly_metric, + labels=write_labels, + ts_future=ts_future, + pred_values=pred_values, + lower_values=lower_values, + upper_values=upper_values, + is_anomaly=is_anomaly, + outside_ratio=outside_ratio, + mean_abs_err=mean_abs_err, + mean_rel_err=mean_rel_err, + event_ts=last_real_ts, + ) + + if not ok: + logger.error("[%s] %s 写入预测数据失败", now_str, query) + continue + + LAST_REAL_TS_WRITTEN[key] = last_real_ts + + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") + future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + last_real_str = datetime.fromtimestamp(last_real_ts).strftime("%H:%M:%S") + origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") + + logger.info( + "[%s] %-40s → %-35s strategy=%s status=%s anomaly=%s period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", + now_str, + query, + pred_metric, + state.strategy, + state.status, + is_anomaly, + state.period, + origin_str, + last_real_str, + data_lag, + len(ts_future), + future_start, + future_end, + ) + + save_state() + + +def main() -> None: + load_state() + + logger.info( + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s", + VM_URL, + HISTORY_MINUTES, + HORIZON_SECONDS, + WRITE_HORIZON_SECONDS, + POLL_INTERVAL, + STATE_FILE, + EXTRA_PREDICT_LABELS["forecast"], + ) + + while True: + run_once() + time.sleep(POLL_INTERVAL) + + +if __name__ == "__main__": + main() \ No newline at end of file From 3609fbae4e2fbb33700de79ee1f7730ac81e366c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 21:39:55 +0800 Subject: [PATCH 27/36] fix --- ai/pridict_v4.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/ai/pridict_v4.py b/ai/pridict_v4.py index 774ad3a..c55f21a 100644 --- a/ai/pridict_v4.py +++ b/ai/pridict_v4.py @@ -120,11 +120,11 @@ "anomaly_metric": "vibration_x_anomaly", "strategy": "phase_band", "abs_threshold": 0.12, - "rel_threshold": 0.35, + "rel_threshold": 0.40, "smooth_window": 5, "band_low_q": 10, "band_high_q": 90, - "band_pad_abs": 0.06, + "band_pad_abs": 0.08, }, { "query": 'vibration_y{device_id="fanuc-cnc"}', @@ -132,11 +132,11 @@ "anomaly_metric": "vibration_y_anomaly", "strategy": "phase_band", "abs_threshold": 0.12, - "rel_threshold": 0.35, + "rel_threshold": 0.40, "smooth_window": 5, "band_low_q": 10, "band_high_q": 90, - "band_pad_abs": 0.06, + "band_pad_abs": 0.08, }, { "query": 'vibration_z{device_id="fanuc-cnc"}', @@ -144,11 +144,11 @@ "anomaly_metric": "vibration_z_anomaly", "strategy": "phase_band", "abs_threshold": 0.12, - "rel_threshold": 0.35, + "rel_threshold": 0.40, "smooth_window": 5, "band_low_q": 10, "band_high_q": 90, - "band_pad_abs": 0.06, + "band_pad_abs": 0.08, }, ] From 1c4217b31d9304cddeb6be16949cc54eaa8d31ec Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 21 May 2026 21:48:41 +0800 Subject: [PATCH 28/36] fix --- ai/pridict_v4.py | 32 ++++++++++++++++---------------- 1 file changed, 16 insertions(+), 16 deletions(-) diff --git a/ai/pridict_v4.py b/ai/pridict_v4.py index c55f21a..8657944 100644 --- a/ai/pridict_v4.py +++ b/ai/pridict_v4.py @@ -119,37 +119,37 @@ "pred_metric": "vibration_x_predicted", "anomaly_metric": "vibration_x_anomaly", "strategy": "phase_band", - "abs_threshold": 0.12, - "rel_threshold": 0.40, + "abs_threshold": 0.18, + "rel_threshold": 0.50, "smooth_window": 5, - "band_low_q": 10, - "band_high_q": 90, - "band_pad_abs": 0.08, + "band_low_q": 2, + "band_high_q": 98, + "band_pad_abs": 0.12, }, { "query": 'vibration_y{device_id="fanuc-cnc"}', "pred_metric": "vibration_y_predicted", "anomaly_metric": "vibration_y_anomaly", "strategy": "phase_band", - "abs_threshold": 0.12, - "rel_threshold": 0.40, + "abs_threshold": 0.18, + "rel_threshold": 0.50, "smooth_window": 5, - "band_low_q": 10, - "band_high_q": 90, - "band_pad_abs": 0.08, + "band_low_q": 2, + "band_high_q": 98, + "band_pad_abs": 0.12, }, { "query": 'vibration_z{device_id="fanuc-cnc"}', "pred_metric": "vibration_z_predicted", "anomaly_metric": "vibration_z_anomaly", "strategy": "phase_band", - "abs_threshold": 0.12, - "rel_threshold": 0.40, + "abs_threshold": 0.18, + "rel_threshold": 0.50, "smooth_window": 5, - "band_low_q": 10, - "band_high_q": 90, - "band_pad_abs": 0.08, - }, + "band_low_q": 2, + "band_high_q": 98, + "band_pad_abs": 0.12, + } ] EXTRA_PREDICT_LABELS = { From 4077e8f416d9d2f9d60df50d551fc817ddacd982 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Fri, 22 May 2026 09:40:02 +0800 Subject: [PATCH 29/36] feat(predict_v5): add predict v5 --- ai/pridict_v5.py | 1794 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 1794 insertions(+) create mode 100644 ai/pridict_v5.py diff --git a/ai/pridict_v5.py b/ai/pridict_v5.py new file mode 100644 index 0000000..6894a66 --- /dev/null +++ b/ai/pridict_v5.py @@ -0,0 +1,1794 @@ +# -*- coding: utf-8 -*- +""" +ProtoForge Predictor v12 + +核心能力： +1. feed_rate / spindle_speed / spindle_current 使用 phase-lock 点预测。 +2. vibration_x / vibration_y / vibration_z 使用 phase-band 预测带。 +3. vibration 类指标： + - predicted 使用平滑后的中位数模板，用于趋势参考。 + - upper/lower 使用原始波动分位数模板 + padding，用于正常波动容忍带。 + - 偶发越界不直接报警，只有持续越界 / 高比例越界 / 严重越界才报警。 +4. 预测起点锚定最后一个真实点 last_real_ts，避免时间错位。 +5. 异常期间冻结健康模板，不学习故障数据。 +6. 故障恢复后等待稳定，再恢复模板学习。 +7. 写入： + - xxx_predicted + - xxx_predicted_upper + - xxx_predicted_lower + - xxx_anomaly + - xxx_anomaly_outside_ratio + - xxx_anomaly_mean_abs_error + - xxx_anomaly_mean_rel_error + - xxx_anomaly_max_consecutive_outside + - xxx_anomaly_max_exceed_ratio +""" + +import json +import logging +import math +import os +import re +import time +from dataclasses import asdict, dataclass +from datetime import datetime, timedelta +from typing import Dict, List, Optional, Tuple + +import numpy as np +import requests + + +# ============================================================================= +# 日志配置 +# ============================================================================= + +logging.basicConfig( + level=logging.INFO, + format="%(asctime)s [%(levelname)s] %(message)s", +) + +logger = logging.getLogger(__name__) + + +# ============================================================================= +# 基础配置 +# ============================================================================= + +VM_URL = "http://localhost:8428" +STATE_FILE = "/tmp/protoforge_predictor_state_v12.json" + +HISTORY_MINUTES = 30 +HORIZON_SECONDS = 120 +POLL_INTERVAL = 30 + +WRITE_HORIZON_SECONDS = min(HORIZON_SECONDS, POLL_INTERVAL) + +QUERY_STEP = "1s" +MIN_POINTS = 120 + +MIN_PERIOD_SECONDS = 5 +MAX_PERIOD_SECONDS = 3600 + +MIN_FULL_CYCLES_FOR_TEMPLATE = 3 +MAX_CYCLES_FOR_TEMPLATE = 8 + +DETECT_WINDOW_SECONDS = 30 +RECOVERY_MIN_SECONDS = 60 + +HEALTHY_EMA_ALPHA = 0.10 +RECOVERY_EMA_ALPHA = 0.25 + +OUTSIDE_RATIO_THRESHOLD = 0.60 +MIN_CONSECUTIVE_OUTSIDE = 5 +SEVERE_EXCEED_RATIO = 1.8 + +VALLEY_QUANTILE = 45 + +MAX_DATA_LAG_SECONDS = 180 + +PHASE_LOCK_MIN_WINDOW_SECONDS = 45 +PHASE_LOCK_MAX_WINDOW_SECONDS = 180 +PHASE_LOCK_PERIOD_SEARCH_RATIO = 0.12 +PHASE_LOCK_ORIGIN_SEARCH_RATIO = 0.35 +PHASE_LOCK_PERIOD_STEP = 1 +PHASE_LOCK_ORIGIN_STEP = 1 + + +# ============================================================================= +# 指标配置 +# ============================================================================= + +PREDICT_TARGETS = [ + { + "query": 'feed_rate{device_id="fanuc-cnc"}', + "pred_metric": "feed_rate_predicted", + "anomaly_metric": "feed_rate_anomaly", + "strategy": "phase_point", + "abs_threshold": 400.0, + "rel_threshold": 0.25, + "smooth_window": 1, + "outside_ratio_threshold": 0.60, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 1.8, + }, + { + "query": 'spindle_speed{device_id="fanuc-cnc"}', + "pred_metric": "spindle_speed_predicted", + "anomaly_metric": "spindle_speed_anomaly", + "strategy": "phase_point", + "abs_threshold": 500.0, + "rel_threshold": 0.25, + "smooth_window": 1, + "outside_ratio_threshold": 0.60, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 1.8, + }, + { + "query": 'spindle_current{device_id="fanuc-cnc"}', + "pred_metric": "spindle_current_predicted", + "anomaly_metric": "spindle_current_anomaly", + "strategy": "phase_point", + "abs_threshold": 5.0, + "rel_threshold": 0.25, + "smooth_window": 1, + "outside_ratio_threshold": 0.60, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 1.8, + }, + { + "query": 'vibration_x{device_id="fanuc-cnc"}', + "pred_metric": "vibration_x_predicted", + "anomaly_metric": "vibration_x_anomaly", + "strategy": "phase_band", + + # vibration 类指标噪声、尖峰较多，不建议用很窄的阈值。 + "abs_threshold": 0.18, + "rel_threshold": 0.55, + + # 平滑只用于相位锁定和 predicted 中位趋势。 + "smooth_window": 5, + + # upper/lower 用原始值分位数，范围放宽，覆盖正常尖峰。 + "band_low_q": 1, + "band_high_q": 99, + "band_pad_abs": 0.15, + + # 偶发越界容忍。 + "outside_ratio_threshold": 0.70, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 2.0, + }, + { + "query": 'vibration_y{device_id="fanuc-cnc"}', + "pred_metric": "vibration_y_predicted", + "anomaly_metric": "vibration_y_anomaly", + "strategy": "phase_band", + "abs_threshold": 0.18, + "rel_threshold": 0.55, + "smooth_window": 5, + "band_low_q": 1, + "band_high_q": 99, + "band_pad_abs": 0.15, + "outside_ratio_threshold": 0.70, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 2.0, + }, + { + "query": 'vibration_z{device_id="fanuc-cnc"}', + "pred_metric": "vibration_z_predicted", + "anomaly_metric": "vibration_z_anomaly", + "strategy": "phase_band", + "abs_threshold": 0.18, + "rel_threshold": 0.55, + "smooth_window": 5, + "band_low_q": 1, + "band_high_q": 99, + "band_pad_abs": 0.15, + "outside_ratio_threshold": 0.70, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 2.0, + }, +] + +EXTRA_PREDICT_LABELS = { + "forecast": "phase_band_health_v12", + "source": "protoforge", +} + +BASELINE_STATUS_HEALTHY = "healthy" +BASELINE_STATUS_ANOMALY = "anomaly" +BASELINE_STATUS_RECOVERING = "recovering" + + +# ============================================================================= +# 状态结构 +# ============================================================================= + +@dataclass +class BaselineState: + period: int + phase_origin_ts: int + template: List[float] + lower_template: List[float] + upper_template: List[float] + strategy: str + status: str + clean_seconds: int + last_update_ts: int + last_seen_ts: int + y_min: float + y_max: float + + +BASELINE_STATES: Dict[str, BaselineState] = {} +LAST_REAL_TS_WRITTEN: Dict[str, int] = {} + + +# ============================================================================= +# VictoriaMetrics 读取 +# ============================================================================= + +def fetch_history(query: str, minutes: int = HISTORY_MINUTES) -> Tuple[List[float], List[float]]: + now = datetime.now() + start = now - timedelta(minutes=minutes) + + try: + resp = requests.get( + f"{VM_URL}/api/v1/query_range", + params={ + "query": query, + "start": start.timestamp(), + "end": now.timestamp(), + "step": QUERY_STEP, + }, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("拉取数据失败 query=%s: %s", query, e) + return [], [] + + try: + result = resp.json().get("data", {}).get("result", []) + except Exception as e: + logger.error("解析 VM 返回失败 query=%s: %s", query, e) + return [], [] + + if not result: + return [], [] + + values = result[0].get("values", []) + + ts = [] + ys = [] + + for item in values: + if len(item) < 2: + continue + + try: + t = float(item[0]) + y = float(item[1]) + except Exception: + continue + + if not math.isfinite(t) or not math.isfinite(y): + continue + + ts.append(t) + ys.append(y) + + return ts, ys + + +def normalize_history(ts: List[float], ys: List[float]) -> Tuple[np.ndarray, np.ndarray]: + if not ts or not ys or len(ts) != len(ys): + return np.array([]), np.array([]) + + data = {} + + for t, y in zip(ts, ys): + try: + sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(sec) or not math.isfinite(val): + continue + + data[sec] = val + + if not data: + return np.array([]), np.array([]) + + sorted_items = sorted(data.items(), key=lambda x: x[0]) + + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) + ys_clean = np.array([x[1] for x in sorted_items], dtype=float) + + if len(ts_clean) < 2: + return ts_clean, ys_clean + + start_sec = int(ts_clean[0]) + end_sec = int(ts_clean[-1]) + + if end_sec <= start_sec: + return ts_clean, ys_clean + + ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) + ys_grid = np.interp(ts_grid, ts_clean, ys_clean) + + return ts_grid, ys_grid + + +# ============================================================================= +# 平滑与预处理 +# ============================================================================= + +def rolling_median(arr: np.ndarray, window: int) -> np.ndarray: + if window <= 1 or len(arr) < window: + return arr.astype(float) + + if window % 2 == 0: + window += 1 + + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + result = [] + + for i in range(len(arr)): + result.append(float(np.median(padded[i:i + window]))) + + return np.array(result, dtype=float) + + +def moving_average(arr: np.ndarray, window: int) -> np.ndarray: + if window <= 1 or len(arr) < window: + return arr.astype(float) + + if window % 2 == 0: + window += 1 + + kernel = np.ones(window, dtype=float) / window + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + return np.convolve(padded, kernel, mode="valid") + + +def preprocess_values(ys_grid: np.ndarray, target: Dict) -> np.ndarray: + strategy = target.get("strategy", "phase_point") + smooth_window = int(target.get("smooth_window", 1)) + + if strategy == "phase_band": + return rolling_median(ys_grid, smooth_window) + + if smooth_window > 1: + return moving_average(ys_grid, smooth_window) + + return ys_grid.astype(float) + + +# ============================================================================= +# 周期估计 +# ============================================================================= + +def estimate_period_by_fft(ys_arr: np.ndarray) -> float: + n = len(ys_arr) + + if n < 8: + return 60.0 + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return 60.0 + + fft_vals = np.fft.rfft(centered) + freqs = np.fft.rfftfreq(n, d=1.0) + + if len(freqs) <= 1: + return 60.0 + + power = np.abs(fft_vals[1:]) + + if len(power) == 0 or np.max(power) <= 0: + return 60.0 + + dominant_idx = int(np.argmax(power)) + 1 + dominant_freq = float(freqs[dominant_idx]) + + if dominant_freq <= 0: + return 60.0 + + period = 1.0 / dominant_freq + + return float(np.clip(period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: + n = len(ys_arr) + + if n < 20: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + centered = ys_arr - np.mean(ys_arr) + + if np.allclose(centered, 0): + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + corr = np.correlate(centered, centered, mode="full")[n - 1:] + + p0 = int(round(init_period)) + left = max(int(MIN_PERIOD_SECONDS), int(max(2, p0 * 0.7))) + right = min(n // 2, int(max(left + 1, p0 * 1.3))) + + if right <= left: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + search = corr[left:right + 1] + + if len(search) == 0: + return float(np.clip(init_period, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + best_lag = left + int(np.argmax(search)) + + return float(np.clip(best_lag, MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS)) + + +def estimate_period_rough(ys_arr: np.ndarray) -> int: + p_fft = estimate_period_by_fft(ys_arr) + p_refined = refine_period_by_autocorr(ys_arr, p_fft) + + period = int(round(p_refined)) + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) + + return int(period) + + +# ============================================================================= +# 谷底检测 +# ============================================================================= + +def find_valley_indices( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + expected_period: int, +) -> List[int]: + n = len(ys_grid) + + if n < max(10, expected_period * 2): + return [] + + period = max(3, int(expected_period)) + smooth_window = max(3, int(round(period * 0.08))) + smooth_window = min(smooth_window, 21) + + ys_smooth = moving_average(ys_grid, smooth_window) + threshold = float(np.percentile(ys_smooth, VALLEY_QUANTILE)) + + candidates = [] + + for i in range(1, n - 1): + if ( + ys_smooth[i] <= ys_smooth[i - 1] + and ys_smooth[i] < ys_smooth[i + 1] + and ys_smooth[i] <= threshold + ): + candidates.append(i) + + if len(candidates) < MIN_FULL_CYCLES_FOR_TEMPLATE: + candidates = [] + + for i in range(1, n - 1): + if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1]: + candidates.append(i) + + if not candidates: + return [] + + min_distance = max(2, int(round(period * 0.55))) + selected = [] + + for idx in candidates: + if not selected: + selected.append(idx) + continue + + if idx - selected[-1] >= min_distance: + selected.append(idx) + continue + + if ys_smooth[idx] < ys_smooth[selected[-1]]: + selected[-1] = idx + + if len(selected) < 2: + return selected + + cleaned = [selected[0]] + + for idx in selected[1:]: + diff = int(ts_grid[idx] - ts_grid[cleaned[-1]]) + + if int(period * 0.55) <= diff <= int(period * 1.60): + cleaned.append(idx) + continue + + if diff < int(period * 0.55): + if ys_smooth[idx] < ys_smooth[cleaned[-1]]: + cleaned[-1] = idx + continue + + cleaned.append(idx) + + return cleaned + + +def detect_period_and_valleys( + ts_grid: np.ndarray, + ys_grid: np.ndarray, +) -> Tuple[int, List[int]]: + rough = estimate_period_rough(ys_grid) + valleys = find_valley_indices(ts_grid, ys_grid, rough) + + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys]) + good = diffs[(diffs >= rough * 0.55) & (diffs <= rough * 1.60)] + + if len(good) > 0: + period = int(round(float(np.median(good)))) + else: + period = rough + else: + period = rough + + period = max(int(MIN_PERIOD_SECONDS), min(int(MAX_PERIOD_SECONDS), period)) + + return int(period), valleys + + +# ============================================================================= +# 模板构建 +# ============================================================================= + +def build_templates_from_valleys( + ts_grid: np.ndarray, + ys_mid_grid: np.ndarray, + ys_band_grid: np.ndarray, + period: int, + valleys: List[int], + target: Dict, +) -> Optional[Tuple[np.ndarray, np.ndarray, np.ndarray]]: + if period <= 1 or len(valleys) < MIN_FULL_CYCLES_FOR_TEMPLATE + 1: + return None + + strategy = target.get("strategy", "phase_point") + low_q = float(target.get("band_low_q", 10)) + high_q = float(target.get("band_high_q", 90)) + + pairs = [] + + for a, b in zip(valleys[:-1], valleys[1:]): + cycle_len = float(ts_grid[b] - ts_grid[a]) + + if period * 0.55 <= cycle_len <= period * 1.60: + pairs.append((a, b, cycle_len)) + + if len(pairs) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + pairs = pairs[-MAX_CYCLES_FOR_TEMPLATE:] + + phase_grid = np.arange(period, dtype=float) + mid_segments = [] + band_segments = [] + weights = [] + + for idx, (a, b, cycle_len) in enumerate(pairs): + seg_ts = ts_grid[a:b + 1] + seg_mid_y = ys_mid_grid[a:b + 1] + seg_band_y = ys_band_grid[a:b + 1] + + if len(seg_mid_y) < 3 or len(seg_band_y) < 3: + continue + + x_old = (seg_ts - seg_ts[0]) / cycle_len * period + + mid_seg = np.interp(phase_grid, x_old, seg_mid_y) + band_seg = np.interp(phase_grid, x_old, seg_band_y) + + mid_segments.append(mid_seg.astype(float)) + band_segments.append(band_seg.astype(float)) + weights.append(0.5 + 0.5 * ((idx + 1) / len(pairs))) + + if len(mid_segments) < MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + mid_arr = np.vstack(mid_segments) + band_arr = np.vstack(band_segments) + w_arr = np.array(weights, dtype=float) + + if strategy == "phase_band": + mid_template = np.percentile(mid_arr, 50, axis=0) + + # upper/lower 使用原始值分布，而不是平滑值分布。 + lower_template = np.percentile(band_arr, low_q, axis=0) + upper_template = np.percentile(band_arr, high_q, axis=0) + else: + mid_template = np.average(mid_arr, axis=0, weights=w_arr) + lower_template = mid_template.copy() + upper_template = mid_template.copy() + + return ( + mid_template.astype(float), + lower_template.astype(float), + upper_template.astype(float), + ) + + +def build_current_baseline( + ts_grid: np.ndarray, + ys_mid_grid: np.ndarray, + ys_band_grid: np.ndarray, + target: Dict, + tail_seconds: Optional[int] = None, +) -> Optional[Tuple[int, int, np.ndarray, np.ndarray, np.ndarray]]: + if len(ys_mid_grid) < MIN_POINTS or len(ys_band_grid) < MIN_POINTS: + return None + + if tail_seconds is not None and tail_seconds > 0: + cutoff = ts_grid[-1] - int(tail_seconds) + mask = ts_grid >= cutoff + ts_use = ts_grid[mask] + ys_mid_use = ys_mid_grid[mask] + ys_band_use = ys_band_grid[mask] + else: + ts_use = ts_grid + ys_mid_use = ys_mid_grid + ys_band_use = ys_band_grid + + if len(ys_mid_use) < MIN_POINTS or len(ys_band_use) < MIN_POINTS: + return None + + period, valleys = detect_period_and_valleys(ts_use, ys_mid_use) + + templates = build_templates_from_valleys( + ts_grid=ts_use, + ys_mid_grid=ys_mid_use, + ys_band_grid=ys_band_use, + period=period, + valleys=valleys, + target=target, + ) + + if templates is None or len(valleys) == 0: + return None + + template, lower_template, upper_template = templates + phase_origin_ts = int(round(float(ts_use[valleys[-1]]))) + + return int(period), phase_origin_ts, template, lower_template, upper_template + + +# ============================================================================= +# 模板预测 +# ============================================================================= + +def circular_template_value(template: np.ndarray, phase: float) -> float: + period = len(template) + + if period == 0: + return 0.0 + + phase = float(phase) % period + i0 = int(math.floor(phase)) % period + i1 = (i0 + 1) % period + frac = phase - math.floor(phase) + + return float((1.0 - frac) * template[i0] + frac * template[i1]) + + +def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: + old_period = len(old_template) + + if old_period == new_period: + return old_template.astype(float) + + if old_period <= 1 or new_period <= 1: + return np.full(new_period, float(np.mean(old_template)), dtype=float) + + old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) + new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) + + old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) + old_y_ext = np.concatenate([old_template, old_template, old_template]) + + return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + + +def predict_template_values( + template: np.ndarray, + period: int, + phase_origin_ts: int, + ts_list: List[int], +) -> np.ndarray: + if period <= 1: + return np.zeros(len(ts_list), dtype=float) + + if len(template) != period: + template = resample_template(template, period) + + values = [] + + for ts in ts_list: + phase = (int(ts) - int(phase_origin_ts)) % period + values.append(circular_template_value(template, phase)) + + return np.array(values, dtype=float) + + +def predict_state_bundle( + state: BaselineState, + ts_list: List[int], +) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: + period = int(state.period) + origin = int(state.phase_origin_ts) + + mid = predict_template_values( + template=np.array(state.template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + lower = predict_template_values( + template=np.array(state.lower_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + upper = predict_template_values( + template=np.array(state.upper_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + + return mid, lower, upper + + +def normalize_origin_near(origin: int, period: int, near_ts: int) -> int: + if period <= 1: + return origin + + origin = int(origin) + period = int(period) + near_ts = int(near_ts) + + while origin + period <= near_ts: + origin += period + + while origin > near_ts: + origin -= period + + return origin + + +def merge_template( + old_template: np.ndarray, + new_template: np.ndarray, + alpha: float, +) -> np.ndarray: + alpha = float(np.clip(alpha, 0.0, 1.0)) + + if len(old_template) != len(new_template): + old_template = resample_template(old_template, len(new_template)) + + merged = (1.0 - alpha) * old_template + alpha * new_template + + return merged.astype(float) + + +# ============================================================================= +# Phase Lock +# ============================================================================= + +def phase_lock_recent( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, +) -> Tuple[int, int, np.ndarray, float]: + base_period = int(state.period) + base_origin = int(state.phase_origin_ts) + base_template = np.array(state.template, dtype=float) + + if base_period <= 1 or len(base_template) <= 1: + ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() + pred = predict_template_values(base_template, base_period, base_origin, ts_recent) + actual = ys_model[-len(ts_recent):].astype(float) + mae = float(np.mean(np.abs(actual - pred))) if len(actual) else 0.0 + return base_period, base_origin, pred, mae + + window_seconds = max( + PHASE_LOCK_MIN_WINDOW_SECONDS, + min(PHASE_LOCK_MAX_WINDOW_SECONDS, int(base_period * 2)), + ) + + cutoff = ts_grid[-1] - window_seconds + mask = ts_grid >= cutoff + + ts_recent_arr = ts_grid[mask].astype(int) + actual = ys_model[mask].astype(float) + + if len(ts_recent_arr) < max(10, DETECT_WINDOW_SECONDS): + ts_recent_arr = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int) + actual = ys_model[-DETECT_WINDOW_SECONDS:].astype(float) + + ts_recent = ts_recent_arr.tolist() + last_ts = int(ts_recent[-1]) + + p_min = max( + int(MIN_PERIOD_SECONDS), + int(round(base_period * (1.0 - PHASE_LOCK_PERIOD_SEARCH_RATIO))), + ) + p_max = min( + int(MAX_PERIOD_SECONDS), + int(round(base_period * (1.0 + PHASE_LOCK_PERIOD_SEARCH_RATIO))), + ) + + best_period = base_period + best_origin = normalize_origin_near(base_origin, base_period, last_ts) + best_template = resample_template(base_template, best_period) + + best_pred = predict_template_values( + template=best_template, + period=best_period, + phase_origin_ts=best_origin, + ts_list=ts_recent, + ) + + best_mae = float(np.mean(np.abs(actual - best_pred))) + + for period in range(p_min, p_max + 1, PHASE_LOCK_PERIOD_STEP): + template = resample_template(base_template, period) + center_origin = normalize_origin_near(base_origin, period, last_ts) + origin_shift = max(2, int(round(period * PHASE_LOCK_ORIGIN_SEARCH_RATIO))) + + for shift in range(-origin_shift, origin_shift + 1, PHASE_LOCK_ORIGIN_STEP): + origin = center_origin + shift + + pred = predict_template_values( + template=template, + period=period, + phase_origin_ts=origin, + ts_list=ts_recent, + ) + + mae = float(np.mean(np.abs(actual - pred))) + penalty = abs(period - base_period) * 0.5 + score = mae + penalty + + best_score = best_mae + abs(best_period - base_period) * 0.5 + + if score < best_score: + best_period = period + best_origin = origin + best_pred = pred + best_mae = mae + + best_origin = normalize_origin_near(best_origin, best_period, last_ts) + + return int(best_period), int(best_origin), best_pred, float(best_mae) + + +# ============================================================================= +# 异常检测 +# ============================================================================= + +def max_consecutive_true(flags: np.ndarray) -> int: + max_count = 0 + current = 0 + + for flag in flags: + if bool(flag): + current += 1 + max_count = max(max_count, current) + else: + current = 0 + + return int(max_count) + + +def calc_point_bounds( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[np.ndarray, np.ndarray]: + threshold = np.maximum(abs_threshold, np.abs(pred) * rel_threshold) + return pred - threshold, pred + threshold + + +def calc_final_bounds( + state: BaselineState, + pred: np.ndarray, + lower_raw: np.ndarray, + upper_raw: np.ndarray, + target: Dict, +) -> Tuple[np.ndarray, np.ndarray]: + strategy = target.get("strategy", "phase_point") + abs_threshold = float(target.get("abs_threshold", 1.0)) + rel_threshold = float(target.get("rel_threshold", 0.25)) + + if strategy == "phase_band": + pad_abs = float(target.get("band_pad_abs", abs_threshold)) + + # 对 vibration 类指标：边界更像正常波动容忍带，不是硬边界。 + dynamic_pad = np.maximum( + pad_abs, + np.abs(pred) * rel_threshold * 0.25, + ) + + lower = lower_raw - dynamic_pad + upper = upper_raw + dynamic_pad + + return lower, upper + + return calc_point_bounds(pred, abs_threshold, rel_threshold) + + +def detect_anomaly( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, +) -> Tuple[bool, float, float, float, int, int, int, float]: + best_period, best_origin, pred_recent, _ = phase_lock_recent( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + ) + + recent_len = len(pred_recent) + + if recent_len <= 0: + return False, 0.0, 0.0, 0.0, best_period, best_origin, 0, 0.0 + + if target.get("strategy", "phase_point") == "phase_band": + actual = ys_actual[-recent_len:].astype(float) + else: + actual = ys_model[-recent_len:].astype(float) + + tmp_state = BaselineState( + period=best_period, + phase_origin_ts=best_origin, + template=state.template, + lower_template=state.lower_template, + upper_template=state.upper_template, + strategy=state.strategy, + status=state.status, + clean_seconds=state.clean_seconds, + last_update_ts=state.last_update_ts, + last_seen_ts=state.last_seen_ts, + y_min=state.y_min, + y_max=state.y_max, + ) + + recent_ts = ts_grid[-recent_len:].astype(int).tolist() + pred, lower_raw, upper_raw = predict_state_bundle(tmp_state, recent_ts) + + lower, upper = calc_final_bounds( + state=tmp_state, + pred=pred, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + above_upper = actual - upper + below_lower = lower - actual + + exceed = np.maximum(above_upper, below_lower) + exceed = np.maximum(exceed, 0.0) + + outside = exceed > 0 + + band_width = np.maximum(upper - lower, 1e-6) + exceed_ratio = exceed / band_width + + abs_err = np.abs(actual - pred) + + outside_ratio = float(np.mean(outside)) + mean_abs_err = float(np.mean(abs_err)) + mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1e-6))) + + max_outside_seconds = max_consecutive_true(outside) + max_exceed_ratio = float(np.max(exceed_ratio)) if len(exceed_ratio) > 0 else 0.0 + + outside_ratio_threshold = float( + target.get("outside_ratio_threshold", OUTSIDE_RATIO_THRESHOLD) + ) + min_consecutive_outside = int( + target.get("min_consecutive_outside", MIN_CONSECUTIVE_OUTSIDE) + ) + severe_exceed_ratio = float( + target.get("severe_exceed_ratio", SEVERE_EXCEED_RATIO) + ) + + # 核心优化： + # 1. 偶发 1~3 个点越界不报警。 + # 2. 持续越界才报警。 + # 3. 高比例越界才报警。 + # 4. 严重越界才立即报警。 + is_anomaly = ( + outside_ratio >= outside_ratio_threshold + or max_outside_seconds >= min_consecutive_outside + or max_exceed_ratio >= severe_exceed_ratio + ) + + return ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + int(best_period), + int(best_origin), + int(max_outside_seconds), + float(max_exceed_ratio), + ) + + +# ============================================================================= +# 状态管理 +# ============================================================================= + +def create_initial_state( + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, + now_sec: int, +) -> Optional[BaselineState]: + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_mid_grid=ys_model, + ys_band_grid=ys_actual, + target=target, + ) + + if baseline is None: + return None + + period, phase_origin_ts, template, lower_template, upper_template = baseline + + return BaselineState( + period=int(period), + phase_origin_ts=int(phase_origin_ts), + template=template.astype(float).tolist(), + lower_template=lower_template.astype(float).tolist(), + upper_template=upper_template.astype(float).tolist(), + strategy=str(target.get("strategy", "phase_point")), + status=BASELINE_STATUS_HEALTHY, + clean_seconds=int(period * MAX_CYCLES_FOR_TEMPLATE), + last_update_ts=now_sec, + last_seen_ts=now_sec, + y_min=float(np.min(ys_actual)), + y_max=float(np.max(ys_actual)), + ) + + +def apply_phase_lock_to_state( + state: BaselineState, + best_period: int, + best_origin: int, +) -> None: + best_period = int(best_period) + + if best_period <= 1: + return + + if len(state.template) != best_period: + state.template = resample_template( + np.array(state.template, dtype=float), + best_period, + ).astype(float).tolist() + + if len(state.lower_template) != best_period: + state.lower_template = resample_template( + np.array(state.lower_template, dtype=float), + best_period, + ).astype(float).tolist() + + if len(state.upper_template) != best_period: + state.upper_template = resample_template( + np.array(state.upper_template, dtype=float), + best_period, + ).astype(float).tolist() + + state.period = best_period + state.phase_origin_ts = int(best_origin) + + +def maybe_update_state( + key: str, + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, +) -> Tuple[Optional[BaselineState], bool, float, float, float, int, float]: + now_sec = int(time.time()) + state = BASELINE_STATES.get(key) + + if state is None: + state = create_initial_state( + ts_grid=ts_grid, + ys_model=ys_model, + ys_actual=ys_actual, + target=target, + now_sec=now_sec, + ) + + if state is None: + return None, False, 0.0, 0.0, 0.0, 0, 0.0 + + BASELINE_STATES[key] = state + + logger.info( + "初始化健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss", + key, + state.strategy, + state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, + ) + + return state, False, 0.0, 0.0, 0.0, 0, 0.0 + + elapsed = max(1, now_sec - int(state.last_seen_ts)) + elapsed = min(elapsed, POLL_INTERVAL * 2) + state.last_seen_ts = now_sec + + ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + best_period, + best_origin, + max_outside_seconds, + max_exceed_ratio, + ) = detect_anomaly( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + ys_actual=ys_actual, + target=target, + ) + + if is_anomaly: + state.status = BASELINE_STATUS_ANOMALY + state.clean_seconds = 0 + BASELINE_STATES[key] = state + + logger.warning( + "检测到异常，冻结模板 key=%s outside_ratio=%.2f max_outside=%ss max_exceed_ratio=%.2f mean_abs_err=%.4f mean_rel_err=%.4f", + key, + outside_ratio, + max_outside_seconds, + max_exceed_ratio, + mean_abs_err, + mean_rel_err, + ) + + return ( + state, + True, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) + + old_period = int(state.period) + old_origin = int(state.phase_origin_ts) + + apply_phase_lock_to_state(state, best_period, best_origin) + + if old_period != state.period or old_origin != state.phase_origin_ts: + logger.info( + "phase-lock key=%s period %s -> %s origin %s -> %s", + key, + old_period, + state.period, + datetime.fromtimestamp(old_origin).strftime("%H:%M:%S"), + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + ) + + if state.status == BASELINE_STATUS_ANOMALY: + state.status = BASELINE_STATUS_RECOVERING + state.clean_seconds = elapsed + BASELINE_STATES[key] = state + + logger.info( + "异常开始恢复 key=%s clean_seconds=%ss", + key, + state.clean_seconds, + ) + + return ( + state, + False, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) + + if state.status == BASELINE_STATUS_RECOVERING: + state.clean_seconds += elapsed + else: + state.status = BASELINE_STATUS_HEALTHY + state.clean_seconds += elapsed + + min_clean_for_update = max( + RECOVERY_MIN_SECONDS, + int(state.period) * MIN_FULL_CYCLES_FOR_TEMPLATE, + ) + + if state.clean_seconds < min_clean_for_update: + BASELINE_STATES[key] = state + return ( + state, + False, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) + + tail_seconds = min( + int(state.clean_seconds), + int(state.period) * MAX_CYCLES_FOR_TEMPLATE, + ) + + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_mid_grid=ys_model, + ys_band_grid=ys_actual, + target=target, + tail_seconds=tail_seconds, + ) + + if baseline is None: + BASELINE_STATES[key] = state + return ( + state, + False, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) + + new_period, new_origin, new_template, new_lower_template, new_upper_template = baseline + + alpha = RECOVERY_EMA_ALPHA if state.status == BASELINE_STATUS_RECOVERING else HEALTHY_EMA_ALPHA + + state.template = merge_template( + np.array(state.template, dtype=float), + new_template, + alpha, + ).astype(float).tolist() + + state.lower_template = merge_template( + np.array(state.lower_template, dtype=float), + new_lower_template, + alpha, + ).astype(float).tolist() + + state.upper_template = merge_template( + np.array(state.upper_template, dtype=float), + new_upper_template, + alpha, + ).astype(float).tolist() + + state.period = int(new_period) + state.phase_origin_ts = int(new_origin) + state.status = BASELINE_STATUS_HEALTHY + state.last_update_ts = now_sec + + if tail_seconds > 0 and len(ys_actual) >= tail_seconds: + state.y_min = float(np.min(ys_actual[-tail_seconds:])) + state.y_max = float(np.max(ys_actual[-tail_seconds:])) + else: + state.y_min = float(np.min(ys_actual)) + state.y_max = float(np.max(ys_actual)) + + BASELINE_STATES[key] = state + + logger.info( + "更新健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss alpha=%.2f", + key, + state.strategy, + state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, + alpha, + ) + + return ( + state, + False, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) + + +# ============================================================================= +# Prometheus 写入 +# ============================================================================= + +def prom_escape_label_value(value: str) -> str: + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) + + +def labels_to_str(labels: Dict[str, str]) -> str: + if not labels: + return "" + + parts = [] + + for k in sorted(labels.keys()): + parts.append(f'{k}="{prom_escape_label_value(labels[k])}"') + + return "{" + ",".join(parts) + "}" + + +def write_series( + metric_name: str, + labels: Dict[str, str], + ts_list: List[int], + values: List[float], +) -> bool: + if not ts_list or not values or len(ts_list) != len(values): + return False + + label_str = labels_to_str(labels) + lines = [] + + for t, y in zip(ts_list, values): + try: + ts_sec = int(round(float(t))) + val = float(y) + except Exception: + continue + + if not math.isfinite(ts_sec) or not math.isfinite(val): + continue + + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_sec * 1000}") + + if not lines: + return False + + payload = "\n".join(lines) + "\n" + + try: + resp = requests.post( + f"{VM_URL}/api/v1/import/prometheus", + data=payload.encode("utf-8"), + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + timeout=10, + ) + resp.raise_for_status() + return True + + except requests.RequestException as e: + logger.error("写入数据失败 metric=%s: %s", metric_name, e) + return False + + +def write_prediction_bundle( + pred_metric: str, + anomaly_metric: str, + labels: Dict[str, str], + ts_future: List[int], + pred_values: np.ndarray, + lower_values: np.ndarray, + upper_values: np.ndarray, + is_anomaly: bool, + outside_ratio: float, + mean_abs_err: float, + mean_rel_err: float, + max_outside_seconds: int, + max_exceed_ratio: float, + event_ts: int, +) -> bool: + ok1 = write_series( + metric_name=pred_metric, + labels=labels, + ts_list=ts_future, + values=pred_values.astype(float).tolist(), + ) + + ok2 = write_series( + metric_name=f"{pred_metric}_lower", + labels=labels, + ts_list=ts_future, + values=lower_values.astype(float).tolist(), + ) + + ok3 = write_series( + metric_name=f"{pred_metric}_upper", + labels=labels, + ts_list=ts_future, + values=upper_values.astype(float).tolist(), + ) + + anomaly_labels = dict(labels) + anomaly_labels["type"] = "prediction_deviation" + + ok4 = write_series( + metric_name=anomaly_metric, + labels=anomaly_labels, + ts_list=[event_ts], + values=[1.0 if is_anomaly else 0.0], + ) + + ok5 = write_series( + metric_name=f"{anomaly_metric}_outside_ratio", + labels=anomaly_labels, + ts_list=[event_ts], + values=[outside_ratio], + ) + + ok6 = write_series( + metric_name=f"{anomaly_metric}_mean_abs_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_abs_err], + ) + + ok7 = write_series( + metric_name=f"{anomaly_metric}_mean_rel_error", + labels=anomaly_labels, + ts_list=[event_ts], + values=[mean_rel_err], + ) + + ok8 = write_series( + metric_name=f"{anomaly_metric}_max_consecutive_outside", + labels=anomaly_labels, + ts_list=[event_ts], + values=[float(max_outside_seconds)], + ) + + ok9 = write_series( + metric_name=f"{anomaly_metric}_max_exceed_ratio", + labels=anomaly_labels, + ts_list=[event_ts], + values=[float(max_exceed_ratio)], + ) + + return ok1 and ok2 and ok3 and ok4 and ok5 and ok6 and ok7 and ok8 and ok9 + + +# ============================================================================= +# 标签解析 +# ============================================================================= + +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) + + +def parse_labels_from_query(query: str) -> Dict[str, str]: + labels = {} + + if "{" not in query or "}" not in query: + return labels + + try: + label_part = query[query.index("{") + 1:query.rindex("}")] + except Exception: + return labels + + for match in _LABEL_PATTERN.finditer(label_part): + key = match.group(1) + value = match.group(2) + + value = ( + value + .replace('\\"', '"') + .replace("\\n", "\n") + .replace("\\\\", "\\") + ) + + labels[key] = value + + return labels + + +def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: + result = {} + + for d in dicts: + if d: + result.update(d) + + return result + + +def series_key(metric_name: str, labels: Dict[str, str]) -> str: + return metric_name + labels_to_str(labels) + + +# ============================================================================= +# 状态持久化 +# ============================================================================= + +def load_state() -> None: + global BASELINE_STATES + + if not os.path.exists(STATE_FILE): + return + + try: + with open(STATE_FILE, "r", encoding="utf-8") as f: + raw = json.load(f) + + states = {} + + for key, value in raw.get("baseline_states", {}).items(): + required_fields = { + "period", + "phase_origin_ts", + "template", + "lower_template", + "upper_template", + "strategy", + "status", + "clean_seconds", + "last_update_ts", + "last_seen_ts", + "y_min", + "y_max", + } + + if not required_fields.issubset(set(value.keys())): + continue + + states[key] = BaselineState(**value) + + BASELINE_STATES = states + + logger.info( + "已加载预测状态文件 %s，状态数量=%d", + STATE_FILE, + len(BASELINE_STATES), + ) + + except Exception as e: + logger.warning("加载预测状态文件失败，将重新学习: %s", e) + + +def save_state() -> None: + try: + raw = { + "baseline_states": { + key: asdict(value) + for key, value in BASELINE_STATES.items() + } + } + + tmp_file = STATE_FILE + ".tmp" + + with open(tmp_file, "w", encoding="utf-8") as f: + json.dump(raw, f, ensure_ascii=False, indent=2) + + os.replace(tmp_file, STATE_FILE) + + except Exception as e: + logger.warning("保存预测状态文件失败: %s", e) + + +# ============================================================================= +# 时间轴 +# ============================================================================= + +def build_prediction_timestamps( + key: str, + last_real_ts: int, + now_sec: int, +) -> Optional[List[int]]: + data_lag = now_sec - last_real_ts + + if data_lag > MAX_DATA_LAG_SECONDS: + logger.warning( + "真实数据延迟过大，跳过预测 key=%s data_lag=%ss max=%ss", + key, + data_lag, + MAX_DATA_LAG_SECONDS, + ) + return None + + last_written_real_ts = LAST_REAL_TS_WRITTEN.get(key) + + if last_written_real_ts is not None and last_real_ts <= int(last_written_real_ts): + logger.info( + "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written_real_ts=%s", + key, + last_real_ts, + last_written_real_ts, + ) + return None + + base_ts = last_real_ts + + return [ + base_ts + i + 1 + for i in range(WRITE_HORIZON_SECONDS) + ] + + +# ============================================================================= +# 主流程 +# ============================================================================= + +def run_once() -> None: + now_str = datetime.now().strftime("%H:%M:%S") + + for target in PREDICT_TARGETS: + query = target["query"] + pred_metric = target["pred_metric"] + anomaly_metric = target["anomaly_metric"] + + ts, ys = fetch_history(query) + + if len(ys) < MIN_POINTS: + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + continue + + ts_grid, ys_grid_raw = normalize_history(ts, ys) + + if len(ys_grid_raw) < MIN_POINTS: + logger.info("[%s] %s 清洗后数据不足（%d 点），跳过", now_str, query, len(ys_grid_raw)) + continue + + ys_grid_model = preprocess_values(ys_grid_raw, target) + + base_labels = parse_labels_from_query(query) + write_labels = merge_labels(base_labels, EXTRA_PREDICT_LABELS) + + key = series_key(pred_metric, write_labels) + + ( + state, + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) = maybe_update_state( + key=key, + ts_grid=ts_grid, + ys_model=ys_grid_model, + ys_actual=ys_grid_raw, + target=target, + ) + + if state is None: + logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + continue + + now_sec = int(time.time()) + last_real_ts = int(ts_grid[-1]) + data_lag = now_sec - last_real_ts + + ts_future = build_prediction_timestamps( + key=key, + last_real_ts=last_real_ts, + now_sec=now_sec, + ) + + if not ts_future: + continue + + pred_values, lower_raw, upper_raw = predict_state_bundle(state, ts_future) + + lower_values, upper_values = calc_final_bounds( + state=state, + pred=pred_values, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + ok = write_prediction_bundle( + pred_metric=pred_metric, + anomaly_metric=anomaly_metric, + labels=write_labels, + ts_future=ts_future, + pred_values=pred_values, + lower_values=lower_values, + upper_values=upper_values, + is_anomaly=is_anomaly, + outside_ratio=outside_ratio, + mean_abs_err=mean_abs_err, + mean_rel_err=mean_rel_err, + max_outside_seconds=max_outside_seconds, + max_exceed_ratio=max_exceed_ratio, + event_ts=last_real_ts, + ) + + if not ok: + logger.error("[%s] %s 写入预测数据失败", now_str, query) + continue + + LAST_REAL_TS_WRITTEN[key] = last_real_ts + + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") + future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + last_real_str = datetime.fromtimestamp(last_real_ts).strftime("%H:%M:%S") + origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") + + logger.info( + "[%s] %-40s → %-35s strategy=%s status=%s anomaly=%s outside=%.2f max_outside=%ss max_exceed=%.2f period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", + now_str, + query, + pred_metric, + state.strategy, + state.status, + is_anomaly, + outside_ratio, + max_outside_seconds, + max_exceed_ratio, + state.period, + origin_str, + last_real_str, + data_lag, + len(ts_future), + future_start, + future_end, + ) + + save_state() + + +def main() -> None: + load_state() + + logger.info( + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s", + VM_URL, + HISTORY_MINUTES, + HORIZON_SECONDS, + WRITE_HORIZON_SECONDS, + POLL_INTERVAL, + STATE_FILE, + EXTRA_PREDICT_LABELS["forecast"], + ) + + while True: + run_once() + time.sleep(POLL_INTERVAL) + + +if __name__ == "__main__": + main() \ No newline at end of file From f9b6506452a75ffd1f1f8beee1ebd1008fdc724d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Mon, 25 May 2026 13:46:29 +0800 Subject: [PATCH 30/36] feat(protoforge): fault update --- protoforge/core/demo.py | 2 +- protoforge/core/fault.py | 187 +++++++++++++++++++++++++++++++++++++ protoforge/models/fault.py | 2 + web/src/views/Devices.vue | 117 ++++++++++++++++++++--- 4 files changed, 296 insertions(+), 12 deletions(-) diff --git a/protoforge/core/demo.py b/protoforge/core/demo.py index ff0b333..b7ccae7 100644 --- a/protoforge/core/demo.py +++ b/protoforge/core/demo.py @@ -108,7 +108,7 @@ async def seed_demo_data(engine: Any, template_manager: Any) -> None: "points": [ {"name": "weight", "address": "net_weight", "data_type": "float32", "generator_type": "random", "min_value": 0.5, "max_value": 50.0}, {"name": "tare", "address": "tare_weight", "data_type": "float32", "generator_type": "fixed", "fixed_value": 2.5}, - {"name": "stable", "address": "stable_flag", "data_type": "bool", "generator_type": "fixed", "fixed_value": true}, + {"name": "stable", "address": "stable_flag", "data_type": "bool", "generator_type": "fixed", "fixed_value": True}, ], }, ] diff --git a/protoforge/core/fault.py b/protoforge/core/fault.py index e72842d..5beba87 100644 --- a/protoforge/core/fault.py +++ b/protoforge/core/fault.py @@ -41,6 +41,7 @@ name="刀具磨损", description="刀具切削刃磨损，切削阻力增大，主轴电流升高，振动增大，进给速率下降", category="mechanical", + scenario_type="trend_drift", default_duration=300.0, tags=["刀具", "磨损", "渐进"], point_faults=[ @@ -67,6 +68,7 @@ name="刀具崩刃", description="刀具突发性崩刃，振动剧烈突增，主轴电流峰值，进给停止", category="mechanical", + scenario_type="sudden_spike", default_duration=15.0, tags=["刀具", "崩刃", "突发"], point_faults=[ @@ -93,6 +95,7 @@ name="主轴过热", description="主轴长时间高负荷运转或冷却不足，电流持续偏高，转速因热保护下降", category="thermal", + scenario_type="trend_drift", default_duration=240.0, tags=["主轴", "过热", "渐进"], point_faults=[ @@ -117,6 +120,7 @@ name="主轴轴承故障", description="主轴轴承磨损或润滑不足，振动幅度持续升高，伴随电流轻微上升", category="mechanical", + scenario_type="trend_drift", default_duration=360.0, tags=["主轴", "轴承", "渐进"], point_faults=[ @@ -141,6 +145,7 @@ name="进给堵转", description="进给轴卡死，进给速率降为零，主轴电流急剧升高", category="process", + scenario_type="sudden_spike", default_duration=20.0, tags=["进给", "堵转", "突发"], point_faults=[ @@ -163,6 +168,7 @@ name="振动异常", description="工件装夹松动或切削共振，三轴振动突然大幅增加", category="mechanical", + scenario_type="sudden_spike", default_duration=60.0, tags=["振动", "装夹", "突发"], point_faults=[ @@ -185,6 +191,7 @@ name="切削液不足", description="切削液供给不足，冷却润滑失效，热量积累导致振动和电流缓慢升高", category="process", + scenario_type="trend_drift", default_duration=480.0, tags=["切削液", "冷却", "渐进"], point_faults=[ @@ -211,6 +218,7 @@ name="电源波动", description="供电电压不稳定，主轴转速和进给速率出现随机波动", category="electrical", + scenario_type="high_noise", default_duration=90.0, tags=["电源", "波动", "突发"], point_faults=[ @@ -222,6 +230,176 @@ multiplier=1.0, noise_scale=150.0), ], ), + + # ================================================================== + # 以下为新增故障类型 + # ================================================================== + + # ------------------------------------------------------------------ + # 传感器强干扰 — 高噪声波动型 + # 场景：电磁干扰、接地不良、信号线屏蔽失效等导致传感器读数剧烈抖动 + # 特征：均值基本不变，但噪声幅度突然增大数倍，信号看起来"毛刺"严重 + # 区别于真实故障：设备本身没有坏，只是采集信号质量变差 + # 模式：瞬间注入，持续期间每次采样都叠加大幅随机噪声 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="sensor_noise", + name="传感器强干扰", + description=( + "【高噪声波动型】电磁干扰或接地不良导致传感器信号质量恶化。" + "均值基本不变，但每次采样叠加大幅随机噪声，曲线呈现密集毛刺。" + "典型场景：变频器附近的传感器、信号线屏蔽层破损、接地回路故障。" + ), + category="electrical", + scenario_type="high_noise", + default_duration=120.0, + tags=["传感器", "干扰", "噪声", "高噪声波动型"], + point_faults=[ + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=8.0), + PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=2.5), + PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=2.5), + PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=3.0), + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=80.0), + ], + ), + + # ------------------------------------------------------------------ + # 换工件/换程序段 — 工况切换型（高速加工 → 低速精加工） + # 场景：CNC 机床切换加工程序，从粗加工切换到精加工 + # 特征：转速降低、进给降低、电流降低，所有指标跳到新的正常范围并稳定 + # 关键：这不是故障！数据本身没有坏，只是工况变了，正常范围完全不同 + # 模式：STEP 阶跃，立即跳到新基线并在整个 duration 内保持 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="mode_switch_fine_machining", + name="切换精加工工况", + description=( + "【工况切换型】从粗加工切换到精加工程序段。" + "主轴转速升高、进给速率降低、切削电流降低，各指标立即跳到新的正常范围并保持稳定。" + "数据本身没有异常，但与粗加工基线相比会触发阈值告警。" + "典型场景：换刀后进入精加工、加工不同特征面、程序跳段。" + ), + category="process", + scenario_type="mode_switch", + default_duration=300.0, + tags=["工况切换", "精加工", "程序段", "工况切换型"], + point_faults=[ + PointFaultConfig(point="spindle_speed", mode=FaultMode.STEP, + multiplier=1.4, noise_scale=30.0), + PointFaultConfig(point="feed_rate", mode=FaultMode.STEP, + multiplier=0.3, noise_scale=10.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.STEP, + multiplier=0.55, noise_scale=0.5), + PointFaultConfig(point="vibration_x", mode=FaultMode.STEP, + multiplier=0.6, noise_scale=0.1), + PointFaultConfig(point="vibration_y", mode=FaultMode.STEP, + multiplier=0.6, noise_scale=0.1), + PointFaultConfig(point="vibration_z", mode=FaultMode.STEP, + multiplier=0.6, noise_scale=0.1), + ], + ), + + # ------------------------------------------------------------------ + # 进入空载工况 — 工况切换型（加工中 → 空载运行） + # 场景：加工完成、等待上料、程序暂停，主轴空转 + # 特征：进给降为 0，电流大幅下降到空载水平，转速维持，振动降低 + # 模式：STEP 阶跃，立即切换到空载基线 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="mode_switch_idle", + name="切换空载工况", + description=( + "【工况切换型】机床进入空载运行状态（加工完成等待上料、程序暂停）。" + "进给速率降为零，主轴电流降至空载水平（约为加工时的 20-30%），" + "主轴转速维持，振动明显降低。" + "典型场景：换料等待、程序暂停、加工间隙、换刀等待。" + ), + category="process", + scenario_type="mode_switch", + default_duration=180.0, + tags=["工况切换", "空载", "等待", "工况切换型"], + point_faults=[ + PointFaultConfig(point="feed_rate", mode=FaultMode.STEP, + target_value=0.0, noise_scale=2.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.STEP, + multiplier=0.22, noise_scale=0.3), + PointFaultConfig(point="vibration_x", mode=FaultMode.STEP, + multiplier=0.25, noise_scale=0.05), + PointFaultConfig(point="vibration_y", mode=FaultMode.STEP, + multiplier=0.25, noise_scale=0.05), + PointFaultConfig(point="vibration_z", mode=FaultMode.STEP, + multiplier=0.25, noise_scale=0.05), + ], + ), + + # ------------------------------------------------------------------ + # 突发电流尖峰 — 突发脉冲型 + # 场景：切削过程中遇到硬质夹杂物、刀具切入角突变、工件材质不均 + # 特征：主轴电流瞬间冲高（持续 2-5 秒），然后恢复正常，其他指标基本不变 + # 区别于刀具崩刃：电流尖峰后能自动恢复，不会导致停机 + # 模式：瞬间注入，持续时间极短 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="current_spike", + name="突发电流尖峰", + description=( + "【突发脉冲型】切削过程中遇到硬质夹杂物或材质不均，主轴电流瞬间冲高后自动恢复。" + "电流短暂升至正常值的 3-4 倍，持续仅数秒，振动轻微抖动，进给基本不受影响。" + "典型场景：铸件内部硬质点、焊缝区域、材料硬度不均匀。" + "与刀具崩刃的区别：能自动恢复，不触发停机报警。" + ), + category="mechanical", + scenario_type="sudden_spike", + default_duration=5.0, + tags=["电流", "尖峰", "脉冲", "突发脉冲型"], + point_faults=[ + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=3.5, noise_scale=1.5), + PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, + multiplier=2.0, noise_scale=0.5), + PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, + multiplier=2.0, noise_scale=0.5), + PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, + multiplier=2.5, noise_scale=0.8), + ], + ), + + # ------------------------------------------------------------------ + # 主轴负载异常 — 关系约束型 + # 场景：刀具钝化但未完全磨损、切削参数不匹配、工件材料变硬 + # 特征：主轴转速正常、进给速率正常，但主轴电流异常升高 + # 关键：单看任何一个指标都"正常"，只有多指标关系才能发现异常 + # 模式：渐进式，电流缓慢爬升，转速和进给保持不变 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="spindle_load_anomaly", + name="主轴负载异常", + description=( + "【关系约束型】主轴转速正常、进给速率正常，但主轴电流异常升高。" + "单看任何一个指标都在正常范围内，只有分析多指标关系才能发现异常。" + "物理含义：切削阻力增大（刀具钝化初期、材料变硬），" + "系统尚未触发保护降速，但电流已超出正常切削功率范围。" + "典型场景：刀具轻度钝化、切削液浓度不足、工件材料批次差异。" + ), + category="mechanical", + scenario_type="relation_constraint", + default_duration=240.0, + tags=["主轴", "负载", "关系约束", "关系约束型"], + point_faults=[ + PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, + multiplier=2.8, noise_scale=1.0), + # 转速和进给保持不变（multiplier=1.0），只叠加极小噪声维持真实感 + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=15.0), + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=5.0), + ], + ), ] # 按 id 索引 @@ -374,6 +552,15 @@ def _compute_value( target = baseline * (1.0 + (pf.multiplier - 1.0) * intensity) else: target = baseline + elif pf.mode == FaultMode.STEP: + # 阶跃模式：立即跳到新基线并在整个 duration 内保持（工况切换专用） + # 与 INSTANT 的区别：STEP 的 multiplier 表示新工况的正常倍数，不受 intensity 缩放 + if pf.target_value is not None: + target = pf.target_value + elif pf.multiplier is not None: + target = baseline * pf.multiplier + else: + target = baseline else: # 渐进模式：随 progress 线性劣化 if pf.target_value is not None: diff --git a/protoforge/models/fault.py b/protoforge/models/fault.py index cc038e0..025da96 100644 --- a/protoforge/models/fault.py +++ b/protoforge/models/fault.py @@ -8,6 +8,7 @@ class FaultMode(str, Enum): """故障注入模式""" INSTANT = "instant" # 瞬间跳变到异常值，持续 duration 后恢复 GRADUAL = "gradual" # 渐进式劣化，随时间线性恶化，到 duration 时达到峰值后恢复 + STEP = "step" # 阶跃切换到新工况基线，整个 duration 内保持新基线（工况切换专用） class FaultStatus(str, Enum): @@ -36,6 +37,7 @@ class FaultTypeDefinition(BaseModel): name: str description: str category: str # 故障分类：mechanical / electrical / thermal / process + scenario_type: str = "trend_drift" # 异常场景类型：trend_drift / sudden_spike / high_noise / mode_switch / relation_constraint default_duration: float = 120.0 # 默认持续时间（秒） point_faults: list[PointFaultConfig] = Field(default_factory=list) tags: list[str] = Field(default_factory=list) diff --git a/web/src/views/Devices.vue b/web/src/views/Devices.vue index 65e0535..36141a4 100644 --- a/web/src/views/Devices.vue +++ b/web/src/views/Devices.vue @@ -92,24 +92,55 @@ - + - 设备：{{ faultTargetDevice?.name }} + 目标设备：{{ faultTargetDevice?.name }} + - -
{{ selectedFaultType.name }} · {{ faultCategoryLabel(selectedFaultType.category) }}
-
{{ selectedFaultType.description }}
-
- 影响测点：{{ selectedFaultType.point_faults.map(p => p.point).join('、') }} + + +
+ + + {{ selectedFaultType.name }} + + {{ scenarioTypeLabel(selectedFaultType.scenario_type) }} + + + {{ faultCategoryLabel(selectedFaultType.category) }} + + + + + {{ selectedFaultType.description }} + + +
+ 影响测点： + + + {{ pf.point }} + + {{ pointFaultModeLabel(pf) }} + + +
- +
+ 秒 + {{ faultIntensityLabel }}（{{ faultIntensity }}） + + · 工况切换型强度不影响切换幅度 + @@ -243,12 +278,17 @@ const columns = [ }, { title: '测点', key: 'points', width: 70, render: (row) => (row.points || []).length }, { - title: '故障', key: 'fault', width: 90, + title: '故障', key: 'fault', width: 130, render: (row) => { const fault = activeFaults.value[row.id] if (!fault || fault.status === 'none') return h(NTag, { size: 'tiny', bordered: false }, () => '正常') const pct = Math.round((fault.progress || 0) * 100) - return h(NTag, { size: 'tiny', type: 'error', bordered: false }, () => `${fault.fault_type_name} ${pct}%`) + const ft = faultTypes.value.find(t => t.id === fault.fault_type_id) + const scenarioLabel = ft ? scenarioTypeLabel(ft.scenario_type) : '' + return h(NSpace, { size: 2, vertical: false, align: 'center' }, () => [ + h(NTag, { size: 'tiny', type: 'error', bordered: false }, () => `${fault.fault_type_name} ${pct}%`), + scenarioLabel ? h(NTag, { size: 'tiny', bordered: false, style: 'font-size:10px;background:#2d1b1b;color:#f87171' }, () => scenarioLabel) : null, + ]) } }, { @@ -351,6 +391,25 @@ const faultTypeOptions = computed(() => faultTypes.value.map(t => ({ label: `${t.name}（${faultCategoryLabel(t.category)}）`, value: t.id })) ) +// 按场景类型分组的故障选项 +const SCENARIO_ORDER = ['trend_drift', 'sudden_spike', 'high_noise', 'mode_switch', 'relation_constraint'] +const faultTypeGroupedOptions = computed(() => { + const groups = {} + for (const t of faultTypes.value) { + const st = t.scenario_type || 'trend_drift' + if (!groups[st]) groups[st] = [] + groups[st].push({ label: t.name, value: t.id }) + } + return SCENARIO_ORDER + .filter(st => groups[st]) + .map(st => ({ + type: 'group', + label: scenarioTypeLabel(st), + key: st, + children: groups[st], + })) +}) + const selectedFaultType = computed(() => faultTypes.value.find(t => t.id === faultTypeId.value) || null ) @@ -368,6 +427,42 @@ function faultCategoryLabel(category) { return map[category] || category } +function scenarioTypeLabel(scenarioType) { + const map = { + trend_drift: '趋势漂移型', + sudden_spike: '突发脉冲型', + high_noise: '高噪声波动型', + mode_switch: '工况切换型', + relation_constraint: '关系约束型', + } + return map[scenarioType] || scenarioType +} + +function scenarioTagType(scenarioType) { + const map = { + trend_drift: 'warning', + sudden_spike: 'error', + high_noise: 'info', + mode_switch: 'success', + relation_constraint: 'default', + } + return map[scenarioType] || 'default' +} + +function pointFaultModeLabel(pf) { + if (pf.mode === 'step') return '→ 阶跃' + if (pf.mode === 'gradual') { + if (pf.multiplier != null) return `→ ×${pf.multiplier}` + if (pf.target_value != null) return `→ ${pf.target_value}` + } + if (pf.mode === 'instant') { + if (pf.target_value != null) return `→ ${pf.target_value}` + if (pf.multiplier != null && pf.multiplier !== 1.0) return `→ ×${pf.multiplier}` + return '± 噪声' + } + return '' +} + function onFaultTypeChange(val) { const t = faultTypes.value.find(f => f.id === val) if (t && t.default_duration) faultDuration.value = t.default_duration From 685ae6b79c28cf8094e20fc6f9f7c00a32d0fe4a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Mon, 25 May 2026 14:00:07 +0800 Subject: [PATCH 31/36] Revert "feat(protoforge): fault update" This reverts commit f9b6506452a75ffd1f1f8beee1ebd1008fdc724d. --- protoforge/core/demo.py | 2 +- protoforge/core/fault.py | 187 ------------------------------------- protoforge/models/fault.py | 2 - web/src/views/Devices.vue | 117 +++-------------------- 4 files changed, 12 insertions(+), 296 deletions(-) diff --git a/protoforge/core/demo.py b/protoforge/core/demo.py index b7ccae7..ff0b333 100644 --- a/protoforge/core/demo.py +++ b/protoforge/core/demo.py @@ -108,7 +108,7 @@ async def seed_demo_data(engine: Any, template_manager: Any) -> None: "points": [ {"name": "weight", "address": "net_weight", "data_type": "float32", "generator_type": "random", "min_value": 0.5, "max_value": 50.0}, {"name": "tare", "address": "tare_weight", "data_type": "float32", "generator_type": "fixed", "fixed_value": 2.5}, - {"name": "stable", "address": "stable_flag", "data_type": "bool", "generator_type": "fixed", "fixed_value": True}, + {"name": "stable", "address": "stable_flag", "data_type": "bool", "generator_type": "fixed", "fixed_value": true}, ], }, ] diff --git a/protoforge/core/fault.py b/protoforge/core/fault.py index 5beba87..e72842d 100644 --- a/protoforge/core/fault.py +++ b/protoforge/core/fault.py @@ -41,7 +41,6 @@ name="刀具磨损", description="刀具切削刃磨损，切削阻力增大，主轴电流升高，振动增大，进给速率下降", category="mechanical", - scenario_type="trend_drift", default_duration=300.0, tags=["刀具", "磨损", "渐进"], point_faults=[ @@ -68,7 +67,6 @@ name="刀具崩刃", description="刀具突发性崩刃，振动剧烈突增，主轴电流峰值，进给停止", category="mechanical", - scenario_type="sudden_spike", default_duration=15.0, tags=["刀具", "崩刃", "突发"], point_faults=[ @@ -95,7 +93,6 @@ name="主轴过热", description="主轴长时间高负荷运转或冷却不足，电流持续偏高，转速因热保护下降", category="thermal", - scenario_type="trend_drift", default_duration=240.0, tags=["主轴", "过热", "渐进"], point_faults=[ @@ -120,7 +117,6 @@ name="主轴轴承故障", description="主轴轴承磨损或润滑不足，振动幅度持续升高，伴随电流轻微上升", category="mechanical", - scenario_type="trend_drift", default_duration=360.0, tags=["主轴", "轴承", "渐进"], point_faults=[ @@ -145,7 +141,6 @@ name="进给堵转", description="进给轴卡死，进给速率降为零，主轴电流急剧升高", category="process", - scenario_type="sudden_spike", default_duration=20.0, tags=["进给", "堵转", "突发"], point_faults=[ @@ -168,7 +163,6 @@ name="振动异常", description="工件装夹松动或切削共振，三轴振动突然大幅增加", category="mechanical", - scenario_type="sudden_spike", default_duration=60.0, tags=["振动", "装夹", "突发"], point_faults=[ @@ -191,7 +185,6 @@ name="切削液不足", description="切削液供给不足，冷却润滑失效，热量积累导致振动和电流缓慢升高", category="process", - scenario_type="trend_drift", default_duration=480.0, tags=["切削液", "冷却", "渐进"], point_faults=[ @@ -218,7 +211,6 @@ name="电源波动", description="供电电压不稳定，主轴转速和进给速率出现随机波动", category="electrical", - scenario_type="high_noise", default_duration=90.0, tags=["电源", "波动", "突发"], point_faults=[ @@ -230,176 +222,6 @@ multiplier=1.0, noise_scale=150.0), ], ), - - # ================================================================== - # 以下为新增故障类型 - # ================================================================== - - # ------------------------------------------------------------------ - # 传感器强干扰 — 高噪声波动型 - # 场景：电磁干扰、接地不良、信号线屏蔽失效等导致传感器读数剧烈抖动 - # 特征：均值基本不变，但噪声幅度突然增大数倍，信号看起来"毛刺"严重 - # 区别于真实故障：设备本身没有坏，只是采集信号质量变差 - # 模式：瞬间注入，持续期间每次采样都叠加大幅随机噪声 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="sensor_noise", - name="传感器强干扰", - description=( - "【高噪声波动型】电磁干扰或接地不良导致传感器信号质量恶化。" - "均值基本不变，但每次采样叠加大幅随机噪声，曲线呈现密集毛刺。" - "典型场景：变频器附近的传感器、信号线屏蔽层破损、接地回路故障。" - ), - category="electrical", - scenario_type="high_noise", - default_duration=120.0, - tags=["传感器", "干扰", "噪声", "高噪声波动型"], - point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=8.0), - PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=2.5), - PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=2.5), - PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=3.0), - PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=80.0), - ], - ), - - # ------------------------------------------------------------------ - # 换工件/换程序段 — 工况切换型（高速加工 → 低速精加工） - # 场景：CNC 机床切换加工程序，从粗加工切换到精加工 - # 特征：转速降低、进给降低、电流降低，所有指标跳到新的正常范围并稳定 - # 关键：这不是故障！数据本身没有坏，只是工况变了，正常范围完全不同 - # 模式：STEP 阶跃，立即跳到新基线并在整个 duration 内保持 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="mode_switch_fine_machining", - name="切换精加工工况", - description=( - "【工况切换型】从粗加工切换到精加工程序段。" - "主轴转速升高、进给速率降低、切削电流降低，各指标立即跳到新的正常范围并保持稳定。" - "数据本身没有异常，但与粗加工基线相比会触发阈值告警。" - "典型场景：换刀后进入精加工、加工不同特征面、程序跳段。" - ), - category="process", - scenario_type="mode_switch", - default_duration=300.0, - tags=["工况切换", "精加工", "程序段", "工况切换型"], - point_faults=[ - PointFaultConfig(point="spindle_speed", mode=FaultMode.STEP, - multiplier=1.4, noise_scale=30.0), - PointFaultConfig(point="feed_rate", mode=FaultMode.STEP, - multiplier=0.3, noise_scale=10.0), - PointFaultConfig(point="spindle_current", mode=FaultMode.STEP, - multiplier=0.55, noise_scale=0.5), - PointFaultConfig(point="vibration_x", mode=FaultMode.STEP, - multiplier=0.6, noise_scale=0.1), - PointFaultConfig(point="vibration_y", mode=FaultMode.STEP, - multiplier=0.6, noise_scale=0.1), - PointFaultConfig(point="vibration_z", mode=FaultMode.STEP, - multiplier=0.6, noise_scale=0.1), - ], - ), - - # ------------------------------------------------------------------ - # 进入空载工况 — 工况切换型（加工中 → 空载运行） - # 场景：加工完成、等待上料、程序暂停，主轴空转 - # 特征：进给降为 0，电流大幅下降到空载水平，转速维持，振动降低 - # 模式：STEP 阶跃，立即切换到空载基线 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="mode_switch_idle", - name="切换空载工况", - description=( - "【工况切换型】机床进入空载运行状态（加工完成等待上料、程序暂停）。" - "进给速率降为零，主轴电流降至空载水平（约为加工时的 20-30%），" - "主轴转速维持，振动明显降低。" - "典型场景：换料等待、程序暂停、加工间隙、换刀等待。" - ), - category="process", - scenario_type="mode_switch", - default_duration=180.0, - tags=["工况切换", "空载", "等待", "工况切换型"], - point_faults=[ - PointFaultConfig(point="feed_rate", mode=FaultMode.STEP, - target_value=0.0, noise_scale=2.0), - PointFaultConfig(point="spindle_current", mode=FaultMode.STEP, - multiplier=0.22, noise_scale=0.3), - PointFaultConfig(point="vibration_x", mode=FaultMode.STEP, - multiplier=0.25, noise_scale=0.05), - PointFaultConfig(point="vibration_y", mode=FaultMode.STEP, - multiplier=0.25, noise_scale=0.05), - PointFaultConfig(point="vibration_z", mode=FaultMode.STEP, - multiplier=0.25, noise_scale=0.05), - ], - ), - - # ------------------------------------------------------------------ - # 突发电流尖峰 — 突发脉冲型 - # 场景：切削过程中遇到硬质夹杂物、刀具切入角突变、工件材质不均 - # 特征：主轴电流瞬间冲高（持续 2-5 秒），然后恢复正常，其他指标基本不变 - # 区别于刀具崩刃：电流尖峰后能自动恢复，不会导致停机 - # 模式：瞬间注入，持续时间极短 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="current_spike", - name="突发电流尖峰", - description=( - "【突发脉冲型】切削过程中遇到硬质夹杂物或材质不均，主轴电流瞬间冲高后自动恢复。" - "电流短暂升至正常值的 3-4 倍，持续仅数秒，振动轻微抖动，进给基本不受影响。" - "典型场景：铸件内部硬质点、焊缝区域、材料硬度不均匀。" - "与刀具崩刃的区别：能自动恢复，不触发停机报警。" - ), - category="mechanical", - scenario_type="sudden_spike", - default_duration=5.0, - tags=["电流", "尖峰", "脉冲", "突发脉冲型"], - point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=3.5, noise_scale=1.5), - PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, - multiplier=2.0, noise_scale=0.5), - PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, - multiplier=2.0, noise_scale=0.5), - PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, - multiplier=2.5, noise_scale=0.8), - ], - ), - - # ------------------------------------------------------------------ - # 主轴负载异常 — 关系约束型 - # 场景：刀具钝化但未完全磨损、切削参数不匹配、工件材料变硬 - # 特征：主轴转速正常、进给速率正常，但主轴电流异常升高 - # 关键：单看任何一个指标都"正常"，只有多指标关系才能发现异常 - # 模式：渐进式，电流缓慢爬升，转速和进给保持不变 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="spindle_load_anomaly", - name="主轴负载异常", - description=( - "【关系约束型】主轴转速正常、进给速率正常，但主轴电流异常升高。" - "单看任何一个指标都在正常范围内，只有分析多指标关系才能发现异常。" - "物理含义：切削阻力增大（刀具钝化初期、材料变硬），" - "系统尚未触发保护降速，但电流已超出正常切削功率范围。" - "典型场景：刀具轻度钝化、切削液浓度不足、工件材料批次差异。" - ), - category="mechanical", - scenario_type="relation_constraint", - default_duration=240.0, - tags=["主轴", "负载", "关系约束", "关系约束型"], - point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, - multiplier=2.8, noise_scale=1.0), - # 转速和进给保持不变（multiplier=1.0），只叠加极小噪声维持真实感 - PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=15.0), - PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=5.0), - ], - ), ] # 按 id 索引 @@ -552,15 +374,6 @@ def _compute_value( target = baseline * (1.0 + (pf.multiplier - 1.0) * intensity) else: target = baseline - elif pf.mode == FaultMode.STEP: - # 阶跃模式：立即跳到新基线并在整个 duration 内保持（工况切换专用） - # 与 INSTANT 的区别：STEP 的 multiplier 表示新工况的正常倍数，不受 intensity 缩放 - if pf.target_value is not None: - target = pf.target_value - elif pf.multiplier is not None: - target = baseline * pf.multiplier - else: - target = baseline else: # 渐进模式：随 progress 线性劣化 if pf.target_value is not None: diff --git a/protoforge/models/fault.py b/protoforge/models/fault.py index 025da96..cc038e0 100644 --- a/protoforge/models/fault.py +++ b/protoforge/models/fault.py @@ -8,7 +8,6 @@ class FaultMode(str, Enum): """故障注入模式""" INSTANT = "instant" # 瞬间跳变到异常值，持续 duration 后恢复 GRADUAL = "gradual" # 渐进式劣化，随时间线性恶化，到 duration 时达到峰值后恢复 - STEP = "step" # 阶跃切换到新工况基线，整个 duration 内保持新基线（工况切换专用） class FaultStatus(str, Enum): @@ -37,7 +36,6 @@ class FaultTypeDefinition(BaseModel): name: str description: str category: str # 故障分类：mechanical / electrical / thermal / process - scenario_type: str = "trend_drift" # 异常场景类型：trend_drift / sudden_spike / high_noise / mode_switch / relation_constraint default_duration: float = 120.0 # 默认持续时间（秒） point_faults: list[PointFaultConfig] = Field(default_factory=list) tags: list[str] = Field(default_factory=list) diff --git a/web/src/views/Devices.vue b/web/src/views/Devices.vue index 36141a4..65e0535 100644 --- a/web/src/views/Devices.vue +++ b/web/src/views/Devices.vue @@ -92,55 +92,24 @@ - + - 目标设备：{{ faultTargetDevice?.name }} - + 设备：{{ faultTargetDevice?.name }} - - -
- - - {{ selectedFaultType.name }} - - {{ scenarioTypeLabel(selectedFaultType.scenario_type) }} - - - {{ faultCategoryLabel(selectedFaultType.category) }} - - - - - {{ selectedFaultType.description }} - - -
- 影响测点： - - - {{ pf.point }} - - {{ pointFaultModeLabel(pf) }} - - - + +
{{ selectedFaultType.name }} · {{ faultCategoryLabel(selectedFaultType.category) }}
+
{{ selectedFaultType.description }}
+
+ 影响测点：{{ selectedFaultType.point_faults.map(p => p.point).join('、') }}
-
- + 秒 - {{ faultIntensityLabel }}（{{ faultIntensity }}） - - · 工况切换型强度不影响切换幅度 - @@ -278,17 +243,12 @@ const columns = [ }, { title: '测点', key: 'points', width: 70, render: (row) => (row.points || []).length }, { - title: '故障', key: 'fault', width: 130, + title: '故障', key: 'fault', width: 90, render: (row) => { const fault = activeFaults.value[row.id] if (!fault || fault.status === 'none') return h(NTag, { size: 'tiny', bordered: false }, () => '正常') const pct = Math.round((fault.progress || 0) * 100) - const ft = faultTypes.value.find(t => t.id === fault.fault_type_id) - const scenarioLabel = ft ? scenarioTypeLabel(ft.scenario_type) : '' - return h(NSpace, { size: 2, vertical: false, align: 'center' }, () => [ - h(NTag, { size: 'tiny', type: 'error', bordered: false }, () => `${fault.fault_type_name} ${pct}%`), - scenarioLabel ? h(NTag, { size: 'tiny', bordered: false, style: 'font-size:10px;background:#2d1b1b;color:#f87171' }, () => scenarioLabel) : null, - ]) + return h(NTag, { size: 'tiny', type: 'error', bordered: false }, () => `${fault.fault_type_name} ${pct}%`) } }, { @@ -391,25 +351,6 @@ const faultTypeOptions = computed(() => faultTypes.value.map(t => ({ label: `${t.name}（${faultCategoryLabel(t.category)}）`, value: t.id })) ) -// 按场景类型分组的故障选项 -const SCENARIO_ORDER = ['trend_drift', 'sudden_spike', 'high_noise', 'mode_switch', 'relation_constraint'] -const faultTypeGroupedOptions = computed(() => { - const groups = {} - for (const t of faultTypes.value) { - const st = t.scenario_type || 'trend_drift' - if (!groups[st]) groups[st] = [] - groups[st].push({ label: t.name, value: t.id }) - } - return SCENARIO_ORDER - .filter(st => groups[st]) - .map(st => ({ - type: 'group', - label: scenarioTypeLabel(st), - key: st, - children: groups[st], - })) -}) - const selectedFaultType = computed(() => faultTypes.value.find(t => t.id === faultTypeId.value) || null ) @@ -427,42 +368,6 @@ function faultCategoryLabel(category) { return map[category] || category } -function scenarioTypeLabel(scenarioType) { - const map = { - trend_drift: '趋势漂移型', - sudden_spike: '突发脉冲型', - high_noise: '高噪声波动型', - mode_switch: '工况切换型', - relation_constraint: '关系约束型', - } - return map[scenarioType] || scenarioType -} - -function scenarioTagType(scenarioType) { - const map = { - trend_drift: 'warning', - sudden_spike: 'error', - high_noise: 'info', - mode_switch: 'success', - relation_constraint: 'default', - } - return map[scenarioType] || 'default' -} - -function pointFaultModeLabel(pf) { - if (pf.mode === 'step') return '→ 阶跃' - if (pf.mode === 'gradual') { - if (pf.multiplier != null) return `→ ×${pf.multiplier}` - if (pf.target_value != null) return `→ ${pf.target_value}` - } - if (pf.mode === 'instant') { - if (pf.target_value != null) return `→ ${pf.target_value}` - if (pf.multiplier != null && pf.multiplier !== 1.0) return `→ ×${pf.multiplier}` - return '± 噪声' - } - return '' -} - function onFaultTypeChange(val) { const t = faultTypes.value.find(f => f.id === val) if (t && t.default_duration) faultDuration.value = t.default_duration From 02174daba0a6b5f0c770243c5f47f082ea2e5d51 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Tue, 26 May 2026 10:47:45 +0800 Subject: [PATCH 32/36] feat(protoforge): update protoforge --- protoforge/core/engine.py | 8 + protoforge/core/fault.py | 290 ++++++++++-------- protoforge/protocols/fanuc/server.py | 79 ++++- protoforge/templates/fanuc/fanuc_0if_cnc.json | 69 ++--- protoforge/templates/modbus/fanuc_cnc.json | 58 +++- web/src/views/Devices.vue | 2 +- 6 files changed, 318 insertions(+), 188 deletions(-) diff --git a/protoforge/core/engine.py b/protoforge/core/engine.py index 059f10e..d8b72aa 100644 --- a/protoforge/core/engine.py +++ b/protoforge/core/engine.py @@ -293,6 +293,14 @@ async def _tick_loop(self) -> None: while self._running: for instance in self._devices.values(): instance.tick() + # 将 DeviceInstance._point_values 同步到协议服务器，保证协议层读到最新值 + server = self._protocol_servers.get(instance.protocol) + if server and hasattr(server, '_behaviors'): + behavior = server._behaviors.get(instance.id) + if behavior is not None: + behavior._values.update(instance._point_values) + if hasattr(behavior, 'sync_from_point_values'): + behavior.sync_from_point_values(instance._point_values) for scenario in self._scenario_instances.values(): scenario.tick() await asyncio.sleep(1.0) diff --git a/protoforge/core/fault.py b/protoforge/core/fault.py index e72842d..11b61a7 100644 --- a/protoforge/core/fault.py +++ b/protoforge/core/fault.py @@ -32,194 +32,228 @@ BUILTIN_FAULT_TYPES: list[FaultTypeDefinition] = [ # ------------------------------------------------------------------ - # 刀具磨损 — 最常见的机加工故障 - # 特征：切削阻力增大 → 主轴电流缓慢爬升，振动幅度增大，进给速率被系统压低 - # 模式：渐进式，持续数分钟，模拟刀具从轻度磨损到需要换刀的过程 - # ------------------------------------------------------------------ - FaultTypeDefinition( - id="tool_wear", - name="刀具磨损", - description="刀具切削刃磨损，切削阻力增大，主轴电流升高，振动增大，进给速率下降", - category="mechanical", - default_duration=300.0, - tags=["刀具", "磨损", "渐进"], - point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, - multiplier=2.2, noise_scale=0.8), - PointFaultConfig(point="vibration_x", mode=FaultMode.GRADUAL, - multiplier=3.0, noise_scale=0.3), - PointFaultConfig(point="vibration_y", mode=FaultMode.GRADUAL, - multiplier=3.0, noise_scale=0.3), - PointFaultConfig(point="vibration_z", mode=FaultMode.GRADUAL, - multiplier=3.5, noise_scale=0.4), - PointFaultConfig(point="feed_rate", mode=FaultMode.GRADUAL, - multiplier=0.45, noise_scale=20.0), - ], - ), - - # ------------------------------------------------------------------ - # 刀具崩刃 — 突发性刀具失效 - # 特征：瞬间冲击 → 振动突增，电流瞬间峰值，进给立即停止 - # 模式：瞬间注入，持续时间短（机床通常会触发报警停机） + # 进给堵转 — 工件夹紧松动或切削量过大导致进给卡死 + # 特征：进给速率瞬间降为0，主轴负载和电流急剧升高，主轴仍在转（区别于崩刃） + # 模式：瞬间注入 # ------------------------------------------------------------------ FaultTypeDefinition( - id="tool_breakage", - name="刀具崩刃", - description="刀具突发性崩刃，振动剧烈突增，主轴电流峰值，进给停止", - category="mechanical", - default_duration=15.0, - tags=["刀具", "崩刃", "突发"], + id="feed_stall", + name="进给堵转", + description="进给轴卡死，进给速率降为零，主轴负载和电流急剧升高，主轴转速维持（区别于崩刃停主轴）", + category="process", + default_duration=20.0, + tags=["进给", "堵转", "突发"], point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=4.5, noise_scale=2.0), - PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, - multiplier=8.0, noise_scale=1.5), - PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, - multiplier=8.0, noise_scale=1.5), - PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, - multiplier=10.0, noise_scale=2.0), PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=2.8, noise_scale=5.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=3.8, noise_scale=1.5), ], ), # ------------------------------------------------------------------ # 主轴过热 — 长时间高负荷或冷却系统故障 - # 特征：主轴电流持续偏高，转速因热保护逐渐降低 + # 特征：主轴负载和电流持续偏高，转速因热保护逐渐降低 # 模式：渐进式，持续时间较长 # ------------------------------------------------------------------ FaultTypeDefinition( id="spindle_overheat", name="主轴过热", - description="主轴长时间高负荷运转或冷却不足，电流持续偏高，转速因热保护下降", + description="主轴长时间高负荷运转或冷却不足，spindle_load和spindle_current持续偏高，转速因热保护渐进下降", category="thermal", default_duration=240.0, tags=["主轴", "过热", "渐进"], point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.GRADUAL, + multiplier=1.6, noise_scale=3.0), PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, multiplier=1.8, noise_scale=1.2), PointFaultConfig(point="spindle_speed", mode=FaultMode.GRADUAL, multiplier=0.6, noise_scale=50.0), - PointFaultConfig(point="vibration_x", mode=FaultMode.GRADUAL, - multiplier=1.5, noise_scale=0.2), - PointFaultConfig(point="vibration_z", mode=FaultMode.GRADUAL, - multiplier=1.5, noise_scale=0.2), ], ), # ------------------------------------------------------------------ - # 主轴轴承故障 — 轴承磨损或润滑不足 - # 特征：振动频率特征变化，整体振动幅度升高，电流略升 - # 模式：渐进式 + # 电源波动 — 供电不稳定 + # 特征：主轴转速和进给速率出现随机波动，电流不稳定 + # 模式：瞬间注入（持续期间持续抖动） + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="power_fluctuation", + name="电源波动", + description="供电电压不稳定，主轴转速和进给速率出现随机波动", + category="electrical", + default_duration=90.0, + tags=["电源", "波动", "突发"], + point_faults=[ + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=300.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=5.0), + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=150.0), + ], + ), + + # ------------------------------------------------------------------ + # 刀具磨损加剧 — 主轴负载趋势漂移 + # 特征：spindle_load 基线随时间缓慢爬升（趋势漂移型），电流同步升高 + # 场景：刀具从轻度磨损到需要换刀的完整过程 + # 模式：渐进式，持续时间长 # ------------------------------------------------------------------ FaultTypeDefinition( - id="spindle_bearing_fault", - name="主轴轴承故障", - description="主轴轴承磨损或润滑不足，振动幅度持续升高，伴随电流轻微上升", - category="mechanical", - default_duration=360.0, - tags=["主轴", "轴承", "渐进"], + id="tool_wear_progressive", + name="刀具磨损加剧", + description="刀具磨损导致切削阻力持续增大，spindle_load基线缓慢爬升至1.8倍，spindle_current同步升高；进给速度由G代码控制不受影响", + category="tool", + default_duration=600.0, + tags=["刀具", "磨损", "负载", "趋势漂移"], point_faults=[ - PointFaultConfig(point="vibration_x", mode=FaultMode.GRADUAL, - multiplier=4.0, noise_scale=0.5), - PointFaultConfig(point="vibration_y", mode=FaultMode.GRADUAL, - multiplier=4.0, noise_scale=0.5), - PointFaultConfig(point="vibration_z", mode=FaultMode.GRADUAL, - multiplier=5.0, noise_scale=0.8), + PointFaultConfig(point="spindle_load", mode=FaultMode.GRADUAL, + multiplier=1.8, noise_scale=3.0), PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, - multiplier=1.3, noise_scale=0.5), + multiplier=1.7, noise_scale=1.5), ], ), # ------------------------------------------------------------------ - # 进给堵转 — 工件夹紧松动或切削量过大导致进给卡死 - # 特征：进给速率瞬间降为 0，主轴电流急剧升高 - # 模式：瞬间注入 + # 刀具崩刃 — 主轴负载突发脉冲 + # 特征：spindle_load 瞬间冲高（可超120%，FANUC最大输出200%），进给停止，CNC停主轴 + # 场景：刀具突发性失效，机床触发过载报警并停机 + # 模式：瞬间注入，持续时间极短 # ------------------------------------------------------------------ FaultTypeDefinition( - id="feed_stall", - name="进给堵转", - description="进给轴卡死，进给速率降为零，主轴电流急剧升高", - category="process", - default_duration=20.0, - tags=["进给", "堵转", "突发"], + id="tool_breakage_sudden", + name="刀具崩刃", + description="刀具突发性崩刃，spindle_load瞬间冲高至正常值3.2倍（可超120%，FANUC最大输出200%），进给停止，CNC触发过载报警并停主轴", + category="tool", + default_duration=10.0, + tags=["刀具", "崩刃", "突发", "过载"], point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=3.2, noise_scale=8.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=4.0, noise_scale=3.0), + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + target_value=0.0, noise_scale=0.0), PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), + PointFaultConfig(point="alarm_status", mode=FaultMode.INSTANT, + target_value=1.0, noise_scale=0.0), + ], + ), + + # ------------------------------------------------------------------ + # 过载保护触发 — 负载/进给反向联动异常（关系约束型） + # 特征：负载超限后CNC自动降进给速率，负载高企与进给降速同时出现 + # 场景：切削参数过激进，CNC自适应保护介入 + # 模式：瞬间注入（持续期间维持异常关系） + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="tool_overload_protection", + name="过载保护触发", + description="主轴负载超限，CNC自动降低进给速率保护刀具，负载高企与进给降速同时出现", + category="tool", + default_duration=120.0, + tags=["刀具", "过载", "进给", "关系约束"], + point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=1.9, noise_scale=4.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=3.8, noise_scale=1.5), - PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, - multiplier=5.0, noise_scale=1.0), + multiplier=1.8, noise_scale=2.0), + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + multiplier=0.35, noise_scale=15.0), ], ), # ------------------------------------------------------------------ - # 振动异常 — 工件装夹松动或共振 - # 特征：三轴振动突然大幅增加，其他指标基本正常 - # 模式：瞬间注入 + # 空切检测 — 刀具未接触工件（工况切换型） + # 特征：spindle_load 跌至空载区间（5-15%），主轴转速和进给速率保持正常 + # 场景：工件装夹偏移、程序坐标错误、工件提前切完 + # 模式：瞬间注入（均值跳变，方差不变） # ------------------------------------------------------------------ FaultTypeDefinition( - id="vibration_spike", - name="振动异常", - description="工件装夹松动或切削共振，三轴振动突然大幅增加", - category="mechanical", - default_duration=60.0, - tags=["振动", "装夹", "突发"], + id="air_cutting", + name="空切检测", + description="刀具未接触工件，spindle_load跌至空载区间(5-15%)，spindle_current降至空转水平，转速进给保持正常", + category="tool", + default_duration=180.0, + tags=["刀具", "空切", "工况切换", "负载"], point_faults=[ - PointFaultConfig(point="vibration_x", mode=FaultMode.INSTANT, - multiplier=6.0, noise_scale=1.0), - PointFaultConfig(point="vibration_y", mode=FaultMode.INSTANT, - multiplier=6.0, noise_scale=1.0), - PointFaultConfig(point="vibration_z", mode=FaultMode.INSTANT, - multiplier=7.0, noise_scale=1.2), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + target_value=8.0, noise_scale=2.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + target_value=2.5, noise_scale=0.3), ], ), # ------------------------------------------------------------------ - # 切削液不足 — 冷却润滑失效 - # 特征：热量积累 → 振动缓慢升高，电流缓慢升高，进给略降 - # 模式：渐进式，速度较慢 + # 积屑瘤 — 切屑粘附刀刃导致周期性负载突刺 + # 特征：spindle_load 在正常基线上出现间歇性冲高后恢复，不是持续爬升 + # 突刺幅度约1.5-2倍基线，持续1-3秒后自行恢复，周期不固定 + # 场景：低速切削、切削液不足、韧性材料（铝合金、不锈钢）加工时常见 + # 模式：瞬间注入（noise_scale 大，模拟随机突刺效果） # ------------------------------------------------------------------ FaultTypeDefinition( - id="coolant_failure", - name="切削液不足", - description="切削液供给不足，冷却润滑失效，热量积累导致振动和电流缓慢升高", - category="process", - default_duration=480.0, - tags=["切削液", "冷却", "渐进"], + id="built_up_edge", + name="积屑瘤", + description="切屑粘附刀刃，spindle_load在正常基线上出现间歇性突刺（1.5-2倍），突刺后自行恢复，区别于磨损的持续爬升", + category="tool", + default_duration=300.0, + tags=["刀具", "积屑瘤", "突刺", "低速切削"], point_faults=[ - PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, - multiplier=1.6, noise_scale=0.8), - PointFaultConfig(point="vibration_x", mode=FaultMode.GRADUAL, - multiplier=2.0, noise_scale=0.3), - PointFaultConfig(point="vibration_y", mode=FaultMode.GRADUAL, - multiplier=2.0, noise_scale=0.3), - PointFaultConfig(point="vibration_z", mode=FaultMode.GRADUAL, - multiplier=2.5, noise_scale=0.4), - PointFaultConfig(point="feed_rate", mode=FaultMode.GRADUAL, - multiplier=0.75, noise_scale=15.0), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=1.7, noise_scale=12.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=1.6, noise_scale=4.0), ], ), # ------------------------------------------------------------------ - # 电源波动 — 供电不稳定 - # 特征：主轴转速和进给速率出现随机波动，电流不稳定 - # 模式：瞬间注入（持续期间持续抖动） + # 刀具涂层剥落 — 负载阶跃后在新基线稳定 + # 特征：spindle_load 出现一次阶跃式跳升（区别于缓慢爬升的磨损）， + # 然后在新的高基线上稳定波动，不会继续爬升也不会恢复 + # 场景：涂层质量问题或切削条件恶劣导致涂层突然失效 + # 模式：瞬间注入（立即跳到新基线，持续维持） # ------------------------------------------------------------------ FaultTypeDefinition( - id="power_fluctuation", - name="电源波动", - description="供电电压不稳定，主轴转速和进给速率出现随机波动", - category="electrical", - default_duration=90.0, - tags=["电源", "波动", "突发"], + id="coating_spalling", + name="刀具涂层剥落", + description="刀具涂层突然失效，spindle_load阶跃式跳升至1.5倍后在新基线稳定波动，区别于磨损的缓慢爬升和崩刃的瞬间冲高", + category="tool", + default_duration=600.0, + tags=["刀具", "涂层", "阶跃", "工况切换"], point_faults=[ - PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=300.0), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=1.5, noise_scale=3.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=5.0), + multiplier=1.4, noise_scale=1.5), + ], + ), + + # ------------------------------------------------------------------ + # 换刀装夹偏移 — 刀具伸出量异常导致负载整体偏高 + # 特征：换刀后 spindle_load 整体偏高（1.4-1.6倍），波动规律正常， + # 不是空切（负载不低），不是磨损（不随时间爬升） + # 场景：刀具伸出量偏长、刀柄锥面未清洁、刀具型号装错 + # 模式：瞬间注入（均值整体偏移，方差不变） + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="tool_offset_error", + name="换刀装夹偏移", + description="换刀后刀具伸出量或装夹位置异常，spindle_load整体偏高(1.4-1.6倍)，波动规律正常，不随时间变化，区别于磨损和空切", + category="tool", + default_duration=3600.0, + tags=["刀具", "装夹", "工况切换", "负载偏移"], + point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + multiplier=1.5, noise_scale=3.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + multiplier=1.4, noise_scale=1.5), PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, - multiplier=1.0, noise_scale=150.0), + multiplier=1.0, noise_scale=5.0), ], ), ] @@ -312,10 +346,10 @@ def apply(self, device: Any) -> None: continue baseline = fault.baseline_values.get(pf.point, 0.0) if baseline == 0.0: - # 基线为 0 时用当前值兜底，避免乘法无效 - try: - baseline = float(device._point_values[pf.point]) or 1.0 - except (TypeError, ValueError): + # 基线为0说明注入时设备处于换刀/停机状态 + # target_value 模式可以直接执行（如崩刃归零、空切归空载） + # multiplier 模式跳过，避免在零基线上产生无意义的值 + if pf.target_value is None: continue device._point_values[pf.point] = self._compute_value( diff --git a/protoforge/protocols/fanuc/server.py b/protoforge/protocols/fanuc/server.py index b0d6f15..e711347 100644 --- a/protoforge/protocols/fanuc/server.py +++ b/protoforge/protocols/fanuc/server.py @@ -22,8 +22,11 @@ def __init__(self, points: list[dict]): "program": "O0001", "speed_override": 100, "feed_override": 100, - "spindle_speed": 3000, - "feed_rate": 500, + "spindle_speed": 0.0, + "feed_rate": 0.0, + "spindle_current": 0.0, + "spindle_load": 0.0, + "tool_number": 1, "absolute_pos": [0.0] * 5, "machine_pos": [0.0] * 5, "relative_pos": [0.0] * 5, @@ -32,6 +35,32 @@ def __init__(self, points: list[dict]): for p in points: self._values[p["name"]] = p.get("fixed_value", 0) + def sync_from_point_values(self, point_values: dict[str, Any]) -> None: + """将 DeviceInstance._point_values 同步到 _cnc_status，保持协议层数据与生成器一致""" + mapping = { + "spindle_speed": "spindle_speed", + "feed_rate": "feed_rate", + "spindle_current": "spindle_current", + "spindle_load": "spindle_load", + "tool_number": "tool_number", + "alarm_status": "alarm", + "run_mode": "mode", + "execution_status": "execution", + "program_name": "program", + "x_absolute": ("absolute_pos", 0), + "y_absolute": ("absolute_pos", 1), + "z_absolute": ("absolute_pos", 2), + } + for point_name, status_key in mapping.items(): + if point_name not in point_values: + continue + val = point_values[point_name] + if isinstance(status_key, tuple): + key, idx = status_key + self._cnc_status[key][idx] = float(val) + else: + self._cnc_status[status_key] = val + async def generate_value(self, point_config: dict[str, Any]) -> Any: name = point_config.get("name", "") return self._values.get(name, 0) @@ -144,6 +173,12 @@ def _process_focas(self, data: bytes) -> bytes | None: return self._handle_cnc_rdspindlespd(req_id) elif func_id == 0x0111: return self._handle_cnc_rdfeed(req_id) + elif func_id == 0x0112: + return self._handle_cnc_rdspload(req_id) + elif func_id == 0x0113: + return self._handle_cnc_rdspmeter(req_id) + elif func_id == 0x0114: + return self._handle_cnc_toolnum(req_id) elif func_id == 0x0120: return self._handle_cnc_alarm(req_id) elif func_id == 0x0130: @@ -247,7 +282,7 @@ def _handle_cnc_rdspindlespd(self, req_id: int) -> bytes: def _handle_cnc_rdfeed(self, req_id: int) -> bytes: behavior = next(iter(self._behaviors.values()), None) - feed = behavior._cnc_status.get("feed_rate", 500) if behavior else 500 + feed = behavior._cnc_status.get("feed_rate", 0.0) if behavior else 0.0 resp = bytearray() resp += struct.pack(" bytes: resp += struct.pack(" bytes: + """cnc_rdspload — 主轴负载率(%)，FANUC FOCAS2 原生接口""" + behavior = next(iter(self._behaviors.values()), None) + load = behavior._cnc_status.get("spindle_load", 0.0) if behavior else 0.0 + + resp = bytearray() + resp += struct.pack(" bytes: + """cnc_rdspmeter — 主轴电流(A)""" + behavior = next(iter(self._behaviors.values()), None) + current = behavior._cnc_status.get("spindle_current", 0.0) if behavior else 0.0 + + resp = bytearray() + resp += struct.pack(" bytes: + """cnc_toolnum — 当前刀号""" + behavior = next(iter(self._behaviors.values()), None) + tool = behavior._cnc_status.get("tool_number", 1) if behavior else 1 + + resp = bytearray() + resp += struct.pack(" bytes: behavior = next(iter(self._behaviors.values()), None) alarm = behavior._cnc_status.get("alarm", 0) if behavior else 0 @@ -311,6 +382,8 @@ async def read_points(self, device_id: str) -> list[PointValue]: config = self._device_configs.get(device_id) if not behavior or not config: return [] + # 将 _point_values 同步到 _cnc_status，保证 FOCAS 协议响应与生成器数据一致 + behavior.sync_from_point_values(behavior._values) now = time.time() return [PointValue(name=p.name, value=behavior.get_value(p.name), timestamp=now) for p in config.points] diff --git a/protoforge/templates/fanuc/fanuc_0if_cnc.json b/protoforge/templates/fanuc/fanuc_0if_cnc.json index 39437f3..0f0e362 100644 --- a/protoforge/templates/fanuc/fanuc_0if_cnc.json +++ b/protoforge/templates/fanuc/fanuc_0if_cnc.json @@ -62,77 +62,68 @@ "address": "spindle_speed", "data_type": "float32", "unit": "RPM", - "description": "主轴转速", + "description": "主轴转速，按加工程序阶梯切换：换刀归零，粗铣2000，半精铣4000，精铣6000", "access": "r", - "generator_type": "sawtooth", - "min_value": 1000, + "generator_type": "script", + "min_value": 0, "max_value": 8000, - "generator_config": {"period": 120} + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 2000\nelif t < 150: target = 0\nelif t < 240: target = 4000\nelif t < 270: target = 0\nelse: target = 6000\nnoise = random.gauss(0, 8) if target > 0 else 0; result = round(max(0, target + noise), 1)" + } }, { "name": "feed_rate", "address": "feed_rate", "data_type": "float32", "unit": "mm/min", - "description": "进给速度", + "description": "进给速度，与加工工步联动：换刀时为0，粗铣800，半精铣500，精铣300", "access": "r", - "generator_type": "sine", - "min_value": 100, + "generator_type": "script", + "min_value": 0, "max_value": 5000, - "generator_config": {"period": 60, "phase": 0.0} + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 800\nelif t < 150: target = 0\nelif t < 240: target = 500\nelif t < 270: target = 0\nelse: target = 300\nnoise = random.gauss(0, 5) if target > 0 else 0; result = round(max(0, target + noise), 1)" + } }, { "name": "spindle_current", "address": "spindle_current", "data_type": "float32", "unit": "A", - "description": "主轴电流", - "access": "r", - "generator_type": "sine", - "min_value": 8.0, - "max_value": 32.0, - "generator_config": {"period": 120, "phase": 0.5} - }, - { - "name": "vibration_x", - "address": "vibration_x", - "data_type": "float32", - "unit": "m/s²", - "description": "X轴振动加速度", + "description": "主轴电流，与工步联动：换刀时伺服保持电流约2.5A，粗铣18-24A，半精铣12-18A，精铣8-13A", "access": "r", "generator_type": "script", - "min_value": 0.1, - "max_value": 2.5, + "min_value": 0.0, + "max_value": 40.0, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.5 + 0.3 * math.sin(2 * math.pi * elapsed / 90); noise = random.uniform(-0.15, 0.15); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base, noise_std = 2.5, 0.3\nelif t < 120: base, noise_std = 21.0, 1.2\nelif t < 150: base, noise_std = 2.5, 0.3\nelif t < 240: base, noise_std = 15.0, 1.2\nelif t < 270: base, noise_std = 2.5, 0.3\nelse: base, noise_std = 10.0, 1.2\nresult = round(max(0, base + random.gauss(0, noise_std)), 2)" } }, { - "name": "vibration_y", - "address": "vibration_y", + "name": "spindle_load", + "address": "spindle_load", "data_type": "float32", - "unit": "m/s²", - "description": "Y轴振动加速度", + "unit": "%", + "description": "主轴负载率(0-100%)，与工步联动：换刀时0%，粗铣45-60%，半精铣35-50%，精铣25-40%", "access": "r", "generator_type": "script", - "min_value": 0.1, - "max_value": 2.5, + "min_value": 0.0, + "max_value": 120.0, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.5 + 0.3 * math.sin(2 * math.pi * elapsed / 75 + 1.0); noise = random.uniform(-0.15, 0.15); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base = 0\nelif t < 120: base = 52.0\nelif t < 150: base = 0\nelif t < 240: base = 42.0\nelif t < 270: base = 0\nelse: base = 32.0\nnoise = random.gauss(0, 2.5) if base > 0 else 0; result = round(max(0, min(120.0, base + noise)), 2)" } }, { - "name": "vibration_z", - "address": "vibration_z", - "data_type": "float32", - "unit": "m/s²", - "description": "Z轴振动加速度", + "name": "tool_number", + "address": "tool_number", + "data_type": "uint16", + "description": "当前刀号，与工步联动：换刀阶段切换，粗铣T01，半精铣T02，精铣T03", "access": "r", "generator_type": "script", - "min_value": 0.1, - "max_value": 3.0, + "min_value": 1, + "max_value": 12, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.7 + 0.4 * math.sin(2 * math.pi * elapsed / 60 + 2.1); noise = random.uniform(-0.2, 0.2); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 120: result = 1\nelif t < 240: result = 2\nelse: result = 3" } }, { diff --git a/protoforge/templates/modbus/fanuc_cnc.json b/protoforge/templates/modbus/fanuc_cnc.json index 43622cf..3a21815 100644 --- a/protoforge/templates/modbus/fanuc_cnc.json +++ b/protoforge/templates/modbus/fanuc_cnc.json @@ -11,36 +11,42 @@ "address": "0", "data_type": "uint16", "unit": "RPM", - "description": "主轴实际转速", + "description": "主轴实际转速，与工步联动：换刀归零，粗铣2000，半精铣4000，精铣6000", "access": "r", - "generator_type": "sawtooth", + "generator_type": "script", "min_value": 0, - "max_value": 12000, - "generator_config": {"period": 180} + "max_value": 8000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 2000\nelif t < 150: target = 0\nelif t < 240: target = 4000\nelif t < 270: target = 0\nelse: target = 6000\nnoise = random.gauss(0, 8) if target > 0 else 0; result = round(max(0, target + noise), 1)" + } }, { "name": "feed_rate", "address": "1", "data_type": "float32", "unit": "mm/min", - "description": "实际进给速度", + "description": "实际进给速度，与工步联动：换刀时0，粗铣800，半精铣500，精铣300", "access": "r", - "generator_type": "sine", - "min_value": 200.0, - "max_value": 3000.0, - "generator_config": {"period": 90, "phase": 1.0} + "generator_type": "script", + "min_value": 0, + "max_value": 5000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 800\nelif t < 150: target = 0\nelif t < 240: target = 500\nelif t < 270: target = 0\nelse: target = 300\nnoise = random.gauss(0, 5) if target > 0 else 0; result = round(max(0, target + noise), 1)" + } }, { "name": "spindle_current", "address": "2", "data_type": "float32", "unit": "A", - "description": "主轴电流", + "description": "主轴电流，与工步联动：换刀伺服保持2.5A，粗铣21A，半精铣15A，精铣10A", "access": "r", - "generator_type": "sine", - "min_value": 8.0, - "max_value": 35.0, - "generator_config": {"period": 120, "phase": 2.0} + "generator_type": "script", + "min_value": 0.0, + "max_value": 40.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base, noise_std = 2.5, 0.3\nelif t < 120: base, noise_std = 21.0, 1.2\nelif t < 150: base, noise_std = 2.5, 0.3\nelif t < 240: base, noise_std = 15.0, 1.2\nelif t < 270: base, noise_std = 2.5, 0.3\nelse: base, noise_std = 10.0, 1.2\nresult = round(max(0, base + random.gauss(0, noise_std)), 2)" + } }, { "name": "vibration_x", @@ -84,6 +90,20 @@ "script": "elapsed = context['elapsed']; base = 0.7 + 0.4 * math.sin(2 * math.pi * elapsed / 58 + 2.8); noise = random.uniform(-0.2, 0.2); result = round(max(0.1, base + noise), 3)" } }, + { + "name": "spindle_load", + "address": "29", + "data_type": "float32", + "unit": "%", + "description": "主轴负载率(0-100%)，与工步联动：换刀时0%，粗铣45-60%，半精铣35-50%，精铣25-40%", + "access": "r", + "generator_type": "script", + "min_value": 0.0, + "max_value": 120.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base = 0\nelif t < 120: base = 52.0\nelif t < 150: base = 0\nelif t < 240: base = 42.0\nelif t < 270: base = 0\nelse: base = 32.0\nnoise = random.gauss(0, 2.5) if base > 0 else 0; result = round(max(0, min(120.0, base + noise)), 2)" + } + }, { "name": "spindle_override", "address": "3", @@ -126,10 +146,14 @@ "name": "tool_no", "address": "7", "data_type": "uint16", - "description": "当前刀具号T", + "description": "当前刀具号，与工步联动：粗铣T1，半精铣T2，精铣T3", "access": "r", - "generator_type": "fixed", - "fixed_value": 3 + "generator_type": "script", + "min_value": 1, + "max_value": 12, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 120: result = 1\nelif t < 240: result = 2\nelse: result = 3" + } }, { "name": "abs_x", diff --git a/web/src/views/Devices.vue b/web/src/views/Devices.vue index 65e0535..3a33068 100644 --- a/web/src/views/Devices.vue +++ b/web/src/views/Devices.vue @@ -364,7 +364,7 @@ const faultIntensityLabel = computed(() => { }) function faultCategoryLabel(category) { - const map = { mechanical: '机械', thermal: '热', electrical: '电气', process: '工艺' } + const map = { mechanical: '机械', thermal: '热', electrical: '电气', process: '工艺', tool: '刀具' } return map[category] || category } From 58b2e3685ec83a1b4520874510da9d7ce0542ef4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 27 May 2026 09:36:02 +0800 Subject: [PATCH 33/36] fix(scene): scene different --- protoforge/templates/modbus/fanuc_cnc.json | 24 +- .../templates/modbus/fanuc_cnc_finish.json | 259 ++++++++++++++++++ .../modbus/fanuc_cnc_semi_finish.json | 259 ++++++++++++++++++ 3 files changed, 529 insertions(+), 13 deletions(-) create mode 100644 protoforge/templates/modbus/fanuc_cnc_finish.json create mode 100644 protoforge/templates/modbus/fanuc_cnc_semi_finish.json diff --git a/protoforge/templates/modbus/fanuc_cnc.json b/protoforge/templates/modbus/fanuc_cnc.json index 3a21815..3ff30df 100644 --- a/protoforge/templates/modbus/fanuc_cnc.json +++ b/protoforge/templates/modbus/fanuc_cnc.json @@ -11,13 +11,13 @@ "address": "0", "data_type": "uint16", "unit": "RPM", - "description": "主轴实际转速，与工步联动：换刀归零，粗铣2000，半精铣4000，精铣6000", + "description": "主轴实际转速，粗加工工位：空闲为0，启动后稳定在约2000RPM", "access": "r", "generator_type": "script", "min_value": 0, "max_value": 8000, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 2000\nelif t < 150: target = 0\nelif t < 240: target = 4000\nelif t < 270: target = 0\nelse: target = 6000\nnoise = random.gauss(0, 8) if target > 0 else 0; result = round(max(0, target + noise), 1)" + "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n target = 0\nelif t < 30:\n target = 2000 * ((t - 20) / 10.0)\nelif t < 155:\n target = 2000\nelif t < 165:\n target = 2000 * (1 - ((t - 155) / 10.0))\nelse:\n target = 0\nnoise = random.gauss(0, 15) if target > 100 else 0\nresult = int(round(max(0, min(8000, target + noise))))" } }, { @@ -25,13 +25,13 @@ "address": "1", "data_type": "float32", "unit": "mm/min", - "description": "实际进给速度，与工步联动：换刀时0，粗铣800，半精铣500，精铣300", + "description": "粗加工实际进给速度：切入阶段中低速，稳定粗加工约800mm/min，空闲和主轴启动阶段为0", "access": "r", "generator_type": "script", "min_value": 0, "max_value": 5000, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: target = 0\nelif t < 120: target = 800\nelif t < 150: target = 0\nelif t < 240: target = 500\nelif t < 270: target = 0\nelse: target = 300\nnoise = random.gauss(0, 5) if target > 0 else 0; result = round(max(0, target + noise), 1)" + "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 30:\n target = 0\nelif t < 40:\n target = 500 + 300 * ((t - 30) / 10.0)\nelif t < 140:\n target = 800\nelif t < 155:\n target = 400\nelse:\n target = 0\nnoise = random.gauss(0, 25) if target > 0 else 0\nresult = round(max(0, min(5000, target + noise)), 1)" } }, { @@ -39,13 +39,13 @@ "address": "2", "data_type": "float32", "unit": "A", - "description": "主轴电流，与工步联动：换刀伺服保持2.5A，粗铣21A，半精铣15A，精铣10A", + "description": "粗加工主轴电流：空闲低电流，启动阶段中等电流，切入瞬间升高，稳定粗加工约19~24A", "access": "r", "generator_type": "script", "min_value": 0.0, "max_value": 40.0, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base, noise_std = 2.5, 0.3\nelif t < 120: base, noise_std = 21.0, 1.2\nelif t < 150: base, noise_std = 2.5, 0.3\nelif t < 240: base, noise_std = 15.0, 1.2\nelif t < 270: base, noise_std = 2.5, 0.3\nelse: base, noise_std = 10.0, 1.2\nresult = round(max(0, base + random.gauss(0, noise_std)), 2)" + "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n base, noise_std = 2.2, 0.25\nelif t < 30:\n base, noise_std = 6.0, 0.8\nelif t < 40:\n base, noise_std = 23.5, 1.5\nelif t < 140:\n base, noise_std = 21.5, 1.3\nelif t < 155:\n base, noise_std = 6.0, 0.8\nelif t < 165:\n base, noise_std = 4.0, 0.5\nelse:\n base, noise_std = 2.2, 0.25\nresult = round(max(0, min(40.0, base + random.gauss(0, noise_std))), 2)" } }, { @@ -95,13 +95,13 @@ "address": "29", "data_type": "float32", "unit": "%", - "description": "主轴负载率(0-100%)，与工步联动：换刀时0%，粗铣45-60%，半精铣35-50%，精铣25-40%", + "description": "粗加工主轴负载率，常态0~100%，短时允许到120%；粗加工稳定阶段约45~65%，切入瞬间可能更高", "access": "r", "generator_type": "script", "min_value": 0.0, "max_value": 120.0, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 30: base = 0\nelif t < 120: base = 52.0\nelif t < 150: base = 0\nelif t < 240: base = 42.0\nelif t < 270: base = 0\nelse: base = 32.0\nnoise = random.gauss(0, 2.5) if base > 0 else 0; result = round(max(0, min(120.0, base + noise)), 2)" + "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n base, noise_std = 0.0, 0.0\nelif t < 30:\n base, noise_std = 8.0, 2.0\nelif t < 40:\n base, noise_std = 65.0, 4.0\nelif t < 140:\n base, noise_std = 55.0, 4.0\nelif t < 155:\n base, noise_std = 8.0, 2.0\nelif t < 165:\n base, noise_std = 3.0, 1.0\nelse:\n base, noise_std = 0.0, 0.0\nnoise = random.gauss(0, noise_std) if noise_std > 0 else 0\nresult = round(max(0, min(120.0, base + noise)), 2)" } }, { @@ -146,14 +146,12 @@ "name": "tool_no", "address": "7", "data_type": "uint16", - "description": "当前刀具号，与工步联动：粗铣T1，半精铣T2，精铣T3", + "description": "当前刀具号，粗加工工位固定使用T1", "access": "r", - "generator_type": "script", + "generator_type": "fixed", "min_value": 1, "max_value": 12, - "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 360\nif t < 120: result = 1\nelif t < 240: result = 2\nelse: result = 3" - } + "fixed_value": 1 }, { "name": "abs_x", diff --git a/protoforge/templates/modbus/fanuc_cnc_finish.json b/protoforge/templates/modbus/fanuc_cnc_finish.json new file mode 100644 index 0000000..1243e59 --- /dev/null +++ b/protoforge/templates/modbus/fanuc_cnc_finish.json @@ -0,0 +1,259 @@ +{ + "id": "modbus_fanuc_cnc_finish", + "name": "Fanuc CNC 精铣工位", + "protocol": "modbus_tcp", + "description": "FANUC Series 0i-MF数控系统，精铣工位：主轴约6000RPM，进给约300mm/min，切深小，表面粗糙度Ra0.8~1.6，要求主轴稳定性高", + "manufacturer": "FANUC", + "model": "0i-MF", + "points": [ + { + "name": "spindle_speed", + "address": "0", + "data_type": "uint16", + "unit": "RPM", + "description": "主轴实际转速，精铣工位：空闲为0，启动后稳定在约6000RPM，精铣对转速稳定性要求高，波动小", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 8000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif t < 12:\n target = 0\nelif t < 28:\n target = 6000 * ((t - 12) / 16.0)\nelif t < 95:\n target = 6000\nelif t < 110:\n target = 6000 * (1 - ((t - 95) / 15.0))\nelse:\n target = 0\nnoise = random.gauss(0, 8) if target > 100 else 0\nresult = int(round(max(0, min(8000, target + noise))))" + } + }, + { + "name": "feed_rate", + "address": "1", + "data_type": "float32", + "unit": "mm/min", + "description": "精铣实际进给速度：切入阶段低速，稳定精铣约300mm/min，精铣进给慢且稳定，波动小", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 5000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif t < 28:\n target = 0\nelif t < 36:\n target = 150 + 150 * ((t - 28) / 8.0)\nelif t < 90:\n target = 300\nelif t < 100:\n target = 150\nelse:\n target = 0\nnoise = random.gauss(0, 6) if target > 0 else 0\nresult = round(max(0, min(5000, target + noise)), 1)" + } + }, + { + "name": "spindle_current", + "address": "2", + "data_type": "float32", + "unit": "A", + "description": "精铣主轴电流：空闲约2A，启动约4A，切入峰值约11A，稳定精铣约7~10A，精铣切深小电流低且稳定", + "access": "r", + "generator_type": "script", + "min_value": 0.0, + "max_value": 40.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif t < 12:\n base, noise_std = 2.2, 0.2\nelif t < 28:\n base, noise_std = 4.0, 0.4\nelif t < 36:\n base, noise_std = 11.0, 0.8\nelif t < 90:\n base, noise_std = 8.5, 0.5\nelif t < 100:\n base, noise_std = 4.0, 0.4\nelif t < 110:\n base, noise_std = 3.0, 0.3\nelse:\n base, noise_std = 2.2, 0.2\nresult = round(max(0, min(40.0, base + random.gauss(0, noise_std))), 2)" + } + }, + { + "name": "vibration_x", + "address": "23", + "data_type": "float32", + "unit": "m/s²", + "description": "X轴振动加速度，精铣切深小但转速高，振动幅值小、频率高", + "access": "r", + "generator_type": "script", + "min_value": 0.02, + "max_value": 1.5, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif 28 <= t < 100:\n base = 0.18 + 0.1 * math.sin(2 * math.pi * elapsed / 28 + 0.8)\nelse:\n base = 0.05\nnoise = random.uniform(-0.04, 0.04)\nresult = round(max(0.02, base + noise), 3)" + } + }, + { + "name": "vibration_y", + "address": "25", + "data_type": "float32", + "unit": "m/s²", + "description": "Y轴振动加速度，精铣切深小但转速高，振动幅值小、频率高", + "access": "r", + "generator_type": "script", + "min_value": 0.02, + "max_value": 1.5, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif 28 <= t < 100:\n base = 0.18 + 0.1 * math.sin(2 * math.pi * elapsed / 22 + 1.8)\nelse:\n base = 0.05\nnoise = random.uniform(-0.04, 0.04)\nresult = round(max(0.02, base + noise), 3)" + } + }, + { + "name": "vibration_z", + "address": "27", + "data_type": "float32", + "unit": "m/s²", + "description": "Z轴振动加速度，精铣切深小但转速高，振动幅值小、频率高", + "access": "r", + "generator_type": "script", + "min_value": 0.02, + "max_value": 1.5, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif 28 <= t < 100:\n base = 0.2 + 0.12 * math.sin(2 * math.pi * elapsed / 18 + 2.8)\nelse:\n base = 0.06\nnoise = random.uniform(-0.05, 0.05)\nresult = round(max(0.02, base + noise), 3)" + } + }, + { + "name": "spindle_load", + "address": "29", + "data_type": "float32", + "unit": "%", + "description": "精铣主轴负载率，稳定阶段约15~28%，切入瞬间约32%，精铣切深小负载低且稳定，短时允许到120%", + "access": "r", + "generator_type": "script", + "min_value": 0.0, + "max_value": 120.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 120\nif t < 12:\n base, noise_std = 0.0, 0.0\nelif t < 28:\n base, noise_std = 4.0, 1.0\nelif t < 36:\n base, noise_std = 32.0, 2.0\nelif t < 90:\n base, noise_std = 22.0, 1.5\nelif t < 100:\n base, noise_std = 4.0, 1.0\nelif t < 110:\n base, noise_std = 1.5, 0.5\nelse:\n base, noise_std = 0.0, 0.0\nnoise = random.gauss(0, noise_std) if noise_std > 0 else 0\nresult = round(max(0, min(120.0, base + noise)), 2)" + } + }, + { + "name": "spindle_override", + "address": "3", + "data_type": "uint16", + "unit": "%", + "description": "主轴倍率", + "access": "r", + "generator_type": "fixed", + "fixed_value": 100 + }, + { + "name": "feed_override", + "address": "4", + "data_type": "uint16", + "unit": "%", + "description": "进给倍率", + "access": "r", + "generator_type": "fixed", + "fixed_value": 100 + }, + { + "name": "running_mode", + "address": "5", + "data_type": "uint16", + "description": "运行模式(1=MDI 2=AUTO 3=JOG 4=EDIT 5=HANDLE)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 2 + }, + { + "name": "exec_status", + "address": "6", + "data_type": "uint16", + "description": "执行状态(0=空闲 1=运行 2=暂停 3=M00停 4=M01停)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 1 + }, + { + "name": "tool_no", + "address": "7", + "data_type": "uint16", + "description": "当前刀具号，精铣工位固定使用T3", + "access": "r", + "generator_type": "fixed", + "min_value": 1, + "max_value": 12, + "fixed_value": 3 + }, + { + "name": "abs_x", + "address": "8", + "data_type": "float32", + "unit": "mm", + "description": "X轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -800.0, + "max_value": 800.0 + }, + { + "name": "abs_y", + "address": "10", + "data_type": "float32", + "unit": "mm", + "description": "Y轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -500.0, + "max_value": 500.0 + }, + { + "name": "abs_z", + "address": "12", + "data_type": "float32", + "unit": "mm", + "description": "Z轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -600.0, + "max_value": 200.0 + }, + { + "name": "mach_x", + "address": "14", + "data_type": "float32", + "unit": "mm", + "description": "X轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -800.0, + "max_value": 800.0 + }, + { + "name": "mach_y", + "address": "16", + "data_type": "float32", + "unit": "mm", + "description": "Y轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -500.0, + "max_value": 500.0 + }, + { + "name": "mach_z", + "address": "18", + "data_type": "float32", + "unit": "mm", + "description": "Z轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -600.0, + "max_value": 200.0 + }, + { + "name": "part_count", + "address": "20", + "data_type": "uint16", + "description": "加工计数", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 99999, + "generator_config": { + "script": "elapsed = context['elapsed']; result = min(int(elapsed / 120), 99999)" + } + }, + { + "name": "cycle_time", + "address": "21", + "data_type": "uint16", + "unit": "s", + "description": "循环时间，精铣单件约120s", + "access": "r", + "generator_type": "fixed", + "fixed_value": 120 + }, + { + "name": "alarm_no", + "address": "22", + "data_type": "uint16", + "description": "报警号(0=无报警)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 0 + } + ], + "protocol_config": { + "slave_id": 8 + }, + "tags": ["CNC", "FANUC", "数控", "机床", "发那科", "精铣"] +} diff --git a/protoforge/templates/modbus/fanuc_cnc_semi_finish.json b/protoforge/templates/modbus/fanuc_cnc_semi_finish.json new file mode 100644 index 0000000..9efc897 --- /dev/null +++ b/protoforge/templates/modbus/fanuc_cnc_semi_finish.json @@ -0,0 +1,259 @@ +{ + "id": "modbus_fanuc_cnc_semi_finish", + "name": "Fanuc CNC 半精铣工位", + "protocol": "modbus_tcp", + "description": "FANUC Series 0i-MF数控系统，半精铣工位：主轴约4000RPM，进给约500mm/min，切深中等，表面粗糙度Ra3.2~6.3", + "manufacturer": "FANUC", + "model": "0i-MF", + "points": [ + { + "name": "spindle_speed", + "address": "0", + "data_type": "uint16", + "unit": "RPM", + "description": "主轴实际转速，半精铣工位：空闲为0，启动后稳定在约4000RPM", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 8000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif t < 15:\n target = 0\nelif t < 28:\n target = 4000 * ((t - 15) / 13.0)\nelif t < 120:\n target = 4000\nelif t < 133:\n target = 4000 * (1 - ((t - 120) / 13.0))\nelse:\n target = 0\nnoise = random.gauss(0, 20) if target > 100 else 0\nresult = int(round(max(0, min(8000, target + noise))))" + } + }, + { + "name": "feed_rate", + "address": "1", + "data_type": "float32", + "unit": "mm/min", + "description": "半精铣实际进给速度：切入阶段低速，稳定半精铣约500mm/min，退刀阶段降速", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 5000, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif t < 28:\n target = 0\nelif t < 38:\n target = 250 + 250 * ((t - 28) / 10.0)\nelif t < 115:\n target = 500\nelif t < 128:\n target = 250\nelse:\n target = 0\nnoise = random.gauss(0, 15) if target > 0 else 0\nresult = round(max(0, min(5000, target + noise)), 1)" + } + }, + { + "name": "spindle_current", + "address": "2", + "data_type": "float32", + "unit": "A", + "description": "半精铣主轴电流：空闲约2A，启动约5A，切入峰值约17A，稳定半精铣约13~16A", + "access": "r", + "generator_type": "script", + "min_value": 0.0, + "max_value": 40.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif t < 15:\n base, noise_std = 2.2, 0.25\nelif t < 28:\n base, noise_std = 5.0, 0.6\nelif t < 38:\n base, noise_std = 17.0, 1.2\nelif t < 115:\n base, noise_std = 14.5, 0.9\nelif t < 128:\n base, noise_std = 5.0, 0.6\nelif t < 138:\n base, noise_std = 3.5, 0.4\nelse:\n base, noise_std = 2.2, 0.25\nresult = round(max(0, min(40.0, base + random.gauss(0, noise_std))), 2)" + } + }, + { + "name": "vibration_x", + "address": "23", + "data_type": "float32", + "unit": "m/s²", + "description": "X轴振动加速度，半精铣切深中等，振动幅值中等", + "access": "r", + "generator_type": "script", + "min_value": 0.05, + "max_value": 2.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif 28 <= t < 128:\n base = 0.35 + 0.2 * math.sin(2 * math.pi * elapsed / 55 + 0.8)\nelse:\n base = 0.08\nnoise = random.uniform(-0.08, 0.08)\nresult = round(max(0.05, base + noise), 3)" + } + }, + { + "name": "vibration_y", + "address": "25", + "data_type": "float32", + "unit": "m/s²", + "description": "Y轴振动加速度，半精铣切深中等，振动幅值中等", + "access": "r", + "generator_type": "script", + "min_value": 0.05, + "max_value": 2.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif 28 <= t < 128:\n base = 0.35 + 0.2 * math.sin(2 * math.pi * elapsed / 45 + 1.8)\nelse:\n base = 0.08\nnoise = random.uniform(-0.08, 0.08)\nresult = round(max(0.05, base + noise), 3)" + } + }, + { + "name": "vibration_z", + "address": "27", + "data_type": "float32", + "unit": "m/s²", + "description": "Z轴振动加速度，半精铣切深中等，振动幅值中等", + "access": "r", + "generator_type": "script", + "min_value": 0.05, + "max_value": 2.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif 28 <= t < 128:\n base = 0.4 + 0.22 * math.sin(2 * math.pi * elapsed / 38 + 2.8)\nelse:\n base = 0.1\nnoise = random.uniform(-0.1, 0.1)\nresult = round(max(0.05, base + noise), 3)" + } + }, + { + "name": "spindle_load", + "address": "29", + "data_type": "float32", + "unit": "%", + "description": "半精铣主轴负载率，稳定阶段约30~45%，切入瞬间约48%，短时允许到120%", + "access": "r", + "generator_type": "script", + "min_value": 0.0, + "max_value": 120.0, + "generator_config": { + "script": "elapsed = context['elapsed']; t = elapsed % 150\nif t < 15:\n base, noise_std = 0.0, 0.0\nelif t < 28:\n base, noise_std = 6.0, 1.5\nelif t < 38:\n base, noise_std = 48.0, 3.0\nelif t < 115:\n base, noise_std = 38.0, 2.5\nelif t < 128:\n base, noise_std = 6.0, 1.5\nelif t < 138:\n base, noise_std = 2.5, 0.8\nelse:\n base, noise_std = 0.0, 0.0\nnoise = random.gauss(0, noise_std) if noise_std > 0 else 0\nresult = round(max(0, min(120.0, base + noise)), 2)" + } + }, + { + "name": "spindle_override", + "address": "3", + "data_type": "uint16", + "unit": "%", + "description": "主轴倍率", + "access": "r", + "generator_type": "fixed", + "fixed_value": 100 + }, + { + "name": "feed_override", + "address": "4", + "data_type": "uint16", + "unit": "%", + "description": "进给倍率", + "access": "r", + "generator_type": "fixed", + "fixed_value": 100 + }, + { + "name": "running_mode", + "address": "5", + "data_type": "uint16", + "description": "运行模式(1=MDI 2=AUTO 3=JOG 4=EDIT 5=HANDLE)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 2 + }, + { + "name": "exec_status", + "address": "6", + "data_type": "uint16", + "description": "执行状态(0=空闲 1=运行 2=暂停 3=M00停 4=M01停)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 1 + }, + { + "name": "tool_no", + "address": "7", + "data_type": "uint16", + "description": "当前刀具号，半精铣工位固定使用T2", + "access": "r", + "generator_type": "fixed", + "min_value": 1, + "max_value": 12, + "fixed_value": 2 + }, + { + "name": "abs_x", + "address": "8", + "data_type": "float32", + "unit": "mm", + "description": "X轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -800.0, + "max_value": 800.0 + }, + { + "name": "abs_y", + "address": "10", + "data_type": "float32", + "unit": "mm", + "description": "Y轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -500.0, + "max_value": 500.0 + }, + { + "name": "abs_z", + "address": "12", + "data_type": "float32", + "unit": "mm", + "description": "Z轴绝对坐标", + "access": "r", + "generator_type": "random", + "min_value": -600.0, + "max_value": 200.0 + }, + { + "name": "mach_x", + "address": "14", + "data_type": "float32", + "unit": "mm", + "description": "X轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -800.0, + "max_value": 800.0 + }, + { + "name": "mach_y", + "address": "16", + "data_type": "float32", + "unit": "mm", + "description": "Y轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -500.0, + "max_value": 500.0 + }, + { + "name": "mach_z", + "address": "18", + "data_type": "float32", + "unit": "mm", + "description": "Z轴机械坐标", + "access": "r", + "generator_type": "random", + "min_value": -600.0, + "max_value": 200.0 + }, + { + "name": "part_count", + "address": "20", + "data_type": "uint16", + "description": "加工计数", + "access": "r", + "generator_type": "script", + "min_value": 0, + "max_value": 99999, + "generator_config": { + "script": "elapsed = context['elapsed']; result = min(int(elapsed / 150), 99999)" + } + }, + { + "name": "cycle_time", + "address": "21", + "data_type": "uint16", + "unit": "s", + "description": "循环时间，半精铣单件约150s", + "access": "r", + "generator_type": "fixed", + "fixed_value": 150 + }, + { + "name": "alarm_no", + "address": "22", + "data_type": "uint16", + "description": "报警号(0=无报警)", + "access": "r", + "generator_type": "fixed", + "fixed_value": 0 + } + ], + "protocol_config": { + "slave_id": 7 + }, + "tags": ["CNC", "FANUC", "数控", "机床", "发那科", "半精铣"] +} From 736baadcb2c8073cff474fcad2e2dd6c6662abb9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Wed, 27 May 2026 13:53:54 +0800 Subject: [PATCH 34/36] fix(scene): update scene cnc --- protoforge/templates/modbus/fanuc_cnc.json | 34 +++++++++++----------- 1 file changed, 17 insertions(+), 17 deletions(-) diff --git a/protoforge/templates/modbus/fanuc_cnc.json b/protoforge/templates/modbus/fanuc_cnc.json index 3ff30df..ebbc8f5 100644 --- a/protoforge/templates/modbus/fanuc_cnc.json +++ b/protoforge/templates/modbus/fanuc_cnc.json @@ -11,13 +11,13 @@ "address": "0", "data_type": "uint16", "unit": "RPM", - "description": "主轴实际转速，粗加工工位：空闲为0，启动后稳定在约2000RPM", + "description": "主轴实际转速，粗加工工位：空闲为0，启动后稳定在约2000RPM，每件节拍有随机差异", "access": "r", "generator_type": "script", "min_value": 0, "max_value": 8000, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n target = 0\nelif t < 30:\n target = 2000 * ((t - 20) / 10.0)\nelif t < 155:\n target = 2000\nelif t < 165:\n target = 2000 * (1 - ((t - 155) / 10.0))\nelse:\n target = 0\nnoise = random.gauss(0, 15) if target > 100 else 0\nresult = int(round(max(0, min(8000, target + noise))))" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'spd_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 1)\n cache[ck] = {'dur': rng.randint(-8, 10), 'spd': rng.uniform(-40, 40)}\noff = cache[ck]\nperiod = 180 + off['dur']; t = elapsed % period\nramp = 8; cut_end = period - 25 - ramp\nif t < 20:\n target = 0\nelif t < 20 + ramp:\n target = (2000 + off['spd']) * ((t - 20) / ramp)\nelif t < cut_end:\n target = 2000 + off['spd']\nelif t < cut_end + ramp:\n target = (2000 + off['spd']) * (1 - (t - cut_end) / ramp)\nelse:\n target = 0\nnoise = random.gauss(0, 12) if target > 100 else 0\nresult = int(round(max(0, min(8000, target + noise))))" } }, { @@ -25,13 +25,13 @@ "address": "1", "data_type": "float32", "unit": "mm/min", - "description": "粗加工实际进给速度：切入阶段中低速，稳定粗加工约800mm/min，空闲和主轴启动阶段为0", + "description": "粗加工实际进给速度：切入爬升，稳定粗铣约800mm/min含拐角减速扰动，退出降速，空闲为0", "access": "r", "generator_type": "script", "min_value": 0, "max_value": 5000, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 30:\n target = 0\nelif t < 40:\n target = 500 + 300 * ((t - 30) / 10.0)\nelif t < 140:\n target = 800\nelif t < 155:\n target = 400\nelse:\n target = 0\nnoise = random.gauss(0, 25) if target > 0 else 0\nresult = round(max(0, min(5000, target + noise)), 1)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'feed_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 2)\n cache[ck] = {'dur': rng.randint(-8, 10), 'feed': rng.uniform(-30, 50)}\noff = cache[ck]\nperiod = 180 + off['dur']; t = elapsed % period\ncut_end = period - 25 - 8\nif t < 30:\n target = 0\nelif t < 40:\n target = 400 + 400 * ((t - 30) / 10.0)\nelif t < cut_end:\n base_feed = 800 + off['feed']\n corner = 80 * math.sin(2 * math.pi * elapsed / 23.7) * max(0, math.sin(2 * math.pi * elapsed / 41.3))\n target = base_feed + corner\nelif t < cut_end + 12:\n target = 350\nelse:\n target = 0\nnoise = random.gauss(0, 18) if target > 0 else 0\nresult = round(max(0, min(5000, target + noise)), 1)" } }, { @@ -39,13 +39,13 @@ "address": "2", "data_type": "float32", "unit": "A", - "description": "粗加工主轴电流：空闲低电流，启动阶段中等电流，切入瞬间升高，稳定粗加工约19~24A", + "description": "粗加工主轴电流：空闲约2A，启动约6A，切入峰值约23A，稳定粗铣约19~24A含细碎波动，退刀降低", "access": "r", "generator_type": "script", "min_value": 0.0, "max_value": 40.0, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n base, noise_std = 2.2, 0.25\nelif t < 30:\n base, noise_std = 6.0, 0.8\nelif t < 40:\n base, noise_std = 23.5, 1.5\nelif t < 140:\n base, noise_std = 21.5, 1.3\nelif t < 155:\n base, noise_std = 6.0, 0.8\nelif t < 165:\n base, noise_std = 4.0, 0.5\nelse:\n base, noise_std = 2.2, 0.25\nresult = round(max(0, min(40.0, base + random.gauss(0, noise_std))), 2)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'cur_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 3)\n cache[ck] = {'dur': rng.randint(-8, 10), 'cur': rng.uniform(-1.5, 2.0)}\noff = cache[ck]\nperiod = 180 + off['dur']; t = elapsed % period\ncut_end = period - 25 - 8\nif t < 20:\n base, noise_std = 2.2, 0.2\nelif t < 20 + 8:\n base, noise_std = 6.0, 0.7\nelif t < 40:\n base, noise_std = 23.5 + off['cur'], 1.8\nelif t < cut_end:\n drift = 0.8 * math.sin(2 * math.pi * elapsed / 37.4) + 0.5 * math.sin(2 * math.pi * elapsed / 19.1)\n base, noise_std = 21.5 + off['cur'] + drift, 1.0\nelif t < cut_end + 12:\n base, noise_std = 6.0, 0.7\nelif t < cut_end + 20:\n base, noise_std = 3.5, 0.4\nelse:\n base, noise_std = 2.2, 0.2\nresult = round(max(0, min(40.0, base + random.gauss(0, noise_std))), 2)" } }, { @@ -53,13 +53,13 @@ "address": "23", "data_type": "float32", "unit": "m/s²", - "description": "X轴振动加速度", + "description": "X轴振动加速度，粗铣切削时约0.4~0.9m/s²，空闲时接近0", "access": "r", "generator_type": "script", - "min_value": 0.1, + "min_value": 0.0, "max_value": 2.5, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.5 + 0.3 * math.sin(2 * math.pi * elapsed / 85 + 0.8); noise = random.uniform(-0.15, 0.15); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'vx_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 4)\n cache[ck] = {'dur': rng.randint(-8, 10)}\nperiod = 180 + cache[ck]['dur']; t = elapsed % period\ncut_end = period - 25 - 8\ncutting = 40 <= t < cut_end\nif cutting:\n base = 0.55 + 0.22 * math.sin(2 * math.pi * elapsed / 85 + 0.8) + 0.1 * math.sin(2 * math.pi * elapsed / 17.3)\n noise = random.uniform(-0.12, 0.12)\nelse:\n base = 0.04\n noise = random.uniform(-0.02, 0.02)\nresult = round(max(0.0, base + noise), 3)" } }, { @@ -67,13 +67,13 @@ "address": "25", "data_type": "float32", "unit": "m/s²", - "description": "Y轴振动加速度", + "description": "Y轴振动加速度，粗铣切削时约0.4~0.9m/s²，空闲时接近0", "access": "r", "generator_type": "script", - "min_value": 0.1, + "min_value": 0.0, "max_value": 2.5, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.5 + 0.3 * math.sin(2 * math.pi * elapsed / 70 + 1.8); noise = random.uniform(-0.15, 0.15); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'vy_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 5)\n cache[ck] = {'dur': rng.randint(-8, 10)}\nperiod = 180 + cache[ck]['dur']; t = elapsed % period\ncut_end = period - 25 - 8\ncutting = 40 <= t < cut_end\nif cutting:\n base = 0.5 + 0.2 * math.sin(2 * math.pi * elapsed / 70 + 1.8) + 0.08 * math.sin(2 * math.pi * elapsed / 13.7)\n noise = random.uniform(-0.1, 0.1)\nelse:\n base = 0.04\n noise = random.uniform(-0.02, 0.02)\nresult = round(max(0.0, base + noise), 3)" } }, { @@ -81,13 +81,13 @@ "address": "27", "data_type": "float32", "unit": "m/s²", - "description": "Z轴振动加速度", + "description": "Z轴振动加速度，粗铣切削时约0.5~1.1m/s²（Z向切深方向幅值略大），空闲时接近0", "access": "r", "generator_type": "script", - "min_value": 0.1, + "min_value": 0.0, "max_value": 3.0, "generator_config": { - "script": "elapsed = context['elapsed']; base = 0.7 + 0.4 * math.sin(2 * math.pi * elapsed / 58 + 2.8); noise = random.uniform(-0.2, 0.2); result = round(max(0.1, base + noise), 3)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'vz_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 6)\n cache[ck] = {'dur': rng.randint(-8, 10)}\nperiod = 180 + cache[ck]['dur']; t = elapsed % period\ncut_end = period - 25 - 8\ncutting = 40 <= t < cut_end\nif cutting:\n base = 0.65 + 0.28 * math.sin(2 * math.pi * elapsed / 58 + 2.8) + 0.12 * math.sin(2 * math.pi * elapsed / 11.2)\n noise = random.uniform(-0.15, 0.15)\nelse:\n base = 0.05\n noise = random.uniform(-0.02, 0.02)\nresult = round(max(0.0, base + noise), 3)" } }, { @@ -95,13 +95,13 @@ "address": "29", "data_type": "float32", "unit": "%", - "description": "粗加工主轴负载率，常态0~100%，短时允许到120%；粗加工稳定阶段约45~65%，切入瞬间可能更高", + "description": "粗加工主轴负载率，稳定粗铣约48~68%含刀路扰动，切入瞬间约65~75%，空闲接近0，短时允许到120%", "access": "r", "generator_type": "script", "min_value": 0.0, "max_value": 120.0, "generator_config": { - "script": "elapsed = context['elapsed']; t = elapsed % 180\nif t < 20:\n base, noise_std = 0.0, 0.0\nelif t < 30:\n base, noise_std = 8.0, 2.0\nelif t < 40:\n base, noise_std = 65.0, 4.0\nelif t < 140:\n base, noise_std = 55.0, 4.0\nelif t < 155:\n base, noise_std = 8.0, 2.0\nelif t < 165:\n base, noise_std = 3.0, 1.0\nelse:\n base, noise_std = 0.0, 0.0\nnoise = random.gauss(0, noise_std) if noise_std > 0 else 0\nresult = round(max(0, min(120.0, base + noise)), 2)" + "script": "elapsed = context['elapsed']\ncycle_idx = int(elapsed // 180)\nck = f'load_{cycle_idx}'\nif ck not in cache:\n rng = random.Random(cycle_idx * 7 + 7)\n cache[ck] = {'dur': rng.randint(-8, 10), 'load': rng.uniform(-4.0, 6.0)}\noff = cache[ck]\nperiod = 180 + off['dur']; t = elapsed % period\ncut_end = period - 25 - 8\nif t < 20:\n base, noise_std = 0.0, 0.0\nelif t < 20 + 8:\n base, noise_std = 6.0, 1.5\nelif t < 40:\n base, noise_std = 68.0 + off['load'], 3.5\nelif t < cut_end:\n drift = 4.5 * math.sin(2 * math.pi * elapsed / 37.4) + 2.5 * math.sin(2 * math.pi * elapsed / 19.1) + 1.5 * math.sin(2 * math.pi * elapsed / 7.3)\n base, noise_std = 56.0 + off['load'] + drift, 2.5\nelif t < cut_end + 12:\n base, noise_std = 6.0, 1.5\nelif t < cut_end + 20:\n base, noise_std = 2.0, 0.8\nelse:\n base, noise_std = 0.0, 0.0\nnoise = random.gauss(0, noise_std) if noise_std > 0 else 0\nresult = round(max(0, min(120.0, base + noise)), 2)" } }, { From c9269967283d9cbe7b3f6b384ead70c7ccf2fab7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Thu, 28 May 2026 13:44:34 +0800 Subject: [PATCH 35/36] fix(fault): update fault --- ai/predictor/README.md | 27 +++ ai/predictor/__init__.py | 30 +++ ai/predictor/anomaly.py | 242 ++++++++++++++++++++ ai/predictor/config.py | 186 ++++++++++++++++ ai/predictor/discovery.py | 84 +++++++ ai/predictor/models.py | 114 ++++++++++ ai/predictor/phase_lock.py | 152 +++++++++++++ ai/predictor/profiling.py | 256 +++++++++++++++++++++ ai/predictor/service.py | 333 ++++++++++++++++++++++++++++ ai/predictor/signal.py | 335 ++++++++++++++++++++++++++++ ai/predictor/state.py | 328 +++++++++++++++++++++++++++ ai/predictor/storage.py | 438 ++++++++++++++++++++++++++++++++++++ ai/predictor/template.py | 384 ++++++++++++++++++++++++++++++++ ai/pridict_v5.py | 442 ++++++++++++++++++++++++++----------- protoforge/core/fault.py | 71 +++++- 15 files changed, 3282 insertions(+), 140 deletions(-) create mode 100644 ai/predictor/README.md create mode 100644 ai/predictor/__init__.py create mode 100644 ai/predictor/anomaly.py create mode 100644 ai/predictor/config.py create mode 100644 ai/predictor/discovery.py create mode 100644 ai/predictor/models.py create mode 100644 ai/predictor/phase_lock.py create mode 100644 ai/predictor/profiling.py create mode 100644 ai/predictor/service.py create mode 100644 ai/predictor/signal.py create mode 100644 ai/predictor/state.py create mode 100644 ai/predictor/storage.py create mode 100644 ai/predictor/template.py diff --git a/ai/predictor/README.md b/ai/predictor/README.md new file mode 100644 index 0000000..8fcc022 --- /dev/null +++ b/ai/predictor/README.md @@ -0,0 +1,27 @@ +```bash +ai/predictor/ + ├── __init__.py # 公开 API：PredictorService, run() + ├── config.py # 所有常量，支持环境变量覆盖 + ├── models.py # BaselineState, MetricProfile 数据类 + ├── discovery.py # VM 设备/指标发现 + ├── signal.py # 纯信号处理：平滑、FFT+自相关周期估计、谷底检测 + ├── template.py # 模板构建、预测、重采样、EMA 融合 + ├── phase_lock.py # Phase-lock 相位对齐 + ├── anomaly.py # 异常检测：边界计算、越界统计、三条件判断 + ├── state.py # 状态机：HEALTHY/ANOMALY/RECOVERING 生命周期 + ├── profiling.py # 自适应配置推断：infer_metric_profile, refresh_targets + ├── storage.py # VM 读写、标签工具、状态持久化 + └── service.py # PredictorService 主类（run_once / run） + + 启动方式： + from ai.predictor import run + run() + # 或 + from ai.predictor import PredictorService + PredictorService(vm_url="http://vm:8428").run() + + 主要改进： + - 全局变量（BASELINE_STATES、LAST_REAL_TS_WRITTEN、_TARGETS_CACHE）全部移入 PredictorService 实例属性 + - IO 与计算完全分离：signal.py、template.py、anomaly.py 均为纯函数，无网络请求 + - 每个模块顶部有职责说明，每个公开函数有完整 docstring +``` diff --git a/ai/predictor/__init__.py b/ai/predictor/__init__.py new file mode 100644 index 0000000..9e45810 --- /dev/null +++ b/ai/predictor/__init__.py @@ -0,0 +1,30 @@ +# -*- coding: utf-8 -*- +""" +predictor +~~~~~~~~~ +ProtoForge 预测服务 package。 + +对外暴露： +- ``PredictorService``：预测服务主类，支持 run() 一键启动 +- ``run()``：便捷入口，使用默认配置启动服务 + +快速启动:: + + from ai.predictor import run + run() + +或自定义配置:: + + from ai.predictor import PredictorService + svc = PredictorService(vm_url="http://vm:8428", poll_interval=60) + svc.run() +""" + +from .service import PredictorService + +__all__ = ["PredictorService", "run"] + + +def run() -> None: + """使用默认配置启动预测服务（一行启动）。""" + PredictorService().run() diff --git a/ai/predictor/anomaly.py b/ai/predictor/anomaly.py new file mode 100644 index 0000000..779d38e --- /dev/null +++ b/ai/predictor/anomaly.py @@ -0,0 +1,242 @@ +# -*- coding: utf-8 -*- +""" +predictor.anomaly +~~~~~~~~~~~~~~~~~ +异常检测：判断当前信号是否偏离健康基线。 + +职责： +- 计算预测边界（phase_point 用对称阈值，phase_band 用分位数带） +- 应用物理上下限兜底（来自 override 文件） +- 统计越界比例、连续越界秒数、最大越界倍数 +- 综合三个条件判断是否触发异常 + +依赖：predictor.phase_lock, predictor.template, predictor.config, predictor.models +""" + +from typing import Dict, Tuple + +import numpy as np + +from . import config +from .models import BaselineState +from .phase_lock import phase_lock_recent +from .template import predict_state_bundle + + +def max_consecutive_true(flags: np.ndarray) -> int: + """ + 计算布尔数组中最长连续 True 的长度。 + + 用于统计最长连续越界秒数，是异常判断的条件之一。 + + Args: + flags: 布尔数组（True 表示该点越界） + + Returns: + 最长连续 True 的长度（整数）。 + """ + max_count = 0 + current = 0 + for flag in flags: + if bool(flag): + current += 1 + max_count = max(max_count, current) + else: + current = 0 + return int(max_count) + + +def calc_point_bounds( + pred: np.ndarray, + abs_threshold: float, + rel_threshold: float, +) -> Tuple[np.ndarray, np.ndarray]: + """ + 计算 phase_point 策略的对称预测边界。 + + 边界宽度 = max(abs_threshold, |pred| * rel_threshold)， + 取两者较大值，保证在小值区域有最小绝对容忍度。 + + Args: + pred: 预测中值数组 + abs_threshold: 绝对误差阈值 + rel_threshold: 相对误差阈值（相对于预测值的比例） + + Returns: + (lower, upper) 边界数组对。 + """ + threshold = np.maximum(abs_threshold, np.abs(pred) * rel_threshold) + return pred - threshold, pred + threshold + + +def calc_final_bounds( + state: BaselineState, + pred: np.ndarray, + lower_raw: np.ndarray, + upper_raw: np.ndarray, + target: Dict, +) -> Tuple[np.ndarray, np.ndarray]: + """ + 计算最终预测边界，综合策略、动态填充和物理上下限。 + + phase_band 策略： + 在分位数模板边界基础上，叠加动态填充（band_pad_abs 和相对填充取较大值）， + 覆盖正常的尖峰波动，避免误报。 + + phase_point 策略： + 直接用对称阈值计算边界。 + + 物理上下限（可选）： + 来自 override 文件的 hard_max / hard_min，对边界做最终 clip。 + + Args: + state: 当前基线状态（提供策略信息） + pred: 预测中值数组 + lower_raw: 模板下界数组（phase_band 为分位数，phase_point 等于 pred） + upper_raw: 模板上界数组 + target: target dict，包含阈值和物理上下限配置 + + Returns: + (lower, upper) 最终边界数组对。 + """ + strategy = target.get("strategy", "phase_point") + abs_threshold = float(target.get("abs_threshold", 1.0)) + rel_threshold = float(target.get("rel_threshold", 0.25)) + + if strategy == "phase_band": + pad_abs = float(target.get("band_pad_abs", abs_threshold)) + # 动态填充：取绝对填充和相对填充（预测值的 25% * rel_threshold）的较大值 + dynamic_pad = np.maximum(pad_abs, np.abs(pred) * rel_threshold * 0.25) + lower = lower_raw - dynamic_pad + upper = upper_raw + dynamic_pad + else: + lower, upper = calc_point_bounds(pred, abs_threshold, rel_threshold) + + # 物理上下限兜底（来自 override 文件，可选） + hard_max = target.get("hard_max") + hard_min = target.get("hard_min") + if hard_max is not None: + upper = np.minimum(upper, float(hard_max)) + if hard_min is not None: + lower = np.maximum(lower, float(hard_min)) + + return lower, upper + + +def detect_anomaly( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, +) -> Tuple[bool, float, float, float, int, int, int, float]: + """ + 检测当前信号是否偏离健康基线，返回完整的诊断指标。 + + 流程： + 1. phase-lock 对齐：在最近窗口内找最优 (period, origin) + 2. 用对齐后的参数预测最近窗口的值 + 3. 计算越界统计量 + 4. 按三个条件判断是否异常： + - 越界比例 >= outside_ratio_threshold + - 连续越界秒数 >= min_consecutive_outside + - 最大越界倍数 >= severe_exceed_ratio（单点严重越界立即报警） + + Args: + state: 当前基线状态 + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_model: 平滑后的信号（phase_point 用于比较） + ys_actual: 原始信号（phase_band 用于比较） + target: target dict，包含阈值配置 + + Returns: + (is_anomaly, outside_ratio, mean_abs_err, mean_rel_err, + best_period, best_origin, max_outside_seconds, max_exceed_ratio) + """ + best_period, best_origin, pred_recent, _ = phase_lock_recent( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + target=target, + ) + + recent_len = len(pred_recent) + if recent_len <= 0: + return False, 0.0, 0.0, 0.0, best_period, best_origin, 0, 0.0 + + # phase_band 用原始信号比较（保留真实波动），phase_point 用平滑信号 + if target.get("strategy", "phase_point") == "phase_band": + actual = ys_actual[-recent_len:].astype(float) + else: + actual = ys_model[-recent_len:].astype(float) + + # 用 phase-lock 后的最优参数重新预测（临时 state，不修改原始 state） + tmp_state = BaselineState( + period=best_period, + phase_origin_ts=best_origin, + template=state.template, + lower_template=state.lower_template, + upper_template=state.upper_template, + strategy=state.strategy, + status=state.status, + clean_seconds=state.clean_seconds, + last_update_ts=state.last_update_ts, + last_seen_ts=state.last_seen_ts, + y_min=state.y_min, + y_max=state.y_max, + ) + + recent_ts = ts_grid[-recent_len:].astype(int).tolist() + pred, lower_raw, upper_raw = predict_state_bundle(tmp_state, recent_ts) + + lower, upper = calc_final_bounds( + state=tmp_state, + pred=pred, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + # 计算越界量（负值表示在边界内，clip 到 0） + above_upper = actual - upper + below_lower = lower - actual + exceed = np.maximum(np.maximum(above_upper, below_lower), 0.0) + outside = exceed > 0 + + band_width = np.maximum(upper - lower, 1e-6) + exceed_ratio = exceed / band_width # 越界量相对于边界宽度的倍数 + + abs_err = np.abs(actual - pred) + outside_ratio = float(np.mean(outside)) + mean_abs_err = float(np.mean(abs_err)) + mean_rel_err = float(np.mean(abs_err / np.maximum(np.abs(pred), 1e-6))) + max_outside_seconds = max_consecutive_true(outside) + max_exceed_ratio = float(np.max(exceed_ratio)) if len(exceed_ratio) > 0 else 0.0 + + # 从 target 读取阈值，允许每个指标独立配置 + outside_ratio_threshold = float( + target.get("outside_ratio_threshold", config.OUTSIDE_RATIO_THRESHOLD) + ) + min_consecutive_outside = int( + target.get("min_consecutive_outside", config.MIN_CONSECUTIVE_OUTSIDE) + ) + severe_exceed_ratio = float( + target.get("severe_exceed_ratio", config.SEVERE_EXCEED_RATIO) + ) + + is_anomaly = ( + outside_ratio >= outside_ratio_threshold + or max_outside_seconds >= min_consecutive_outside + or max_exceed_ratio >= severe_exceed_ratio + ) + + return ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + int(best_period), + int(best_origin), + int(max_outside_seconds), + float(max_exceed_ratio), + ) diff --git a/ai/predictor/config.py b/ai/predictor/config.py new file mode 100644 index 0000000..ed5c11b --- /dev/null +++ b/ai/predictor/config.py @@ -0,0 +1,186 @@ +# -*- coding: utf-8 -*- +""" +predictor.config +~~~~~~~~~~~~~~~~ +所有运行时配置常量，集中在此处管理。 + +大部分参数支持通过环境变量覆盖，方便容器化部署时无需修改代码。 +环境变量前缀统一为 ``PROTOFORGE_``。 +""" + +import os +from typing import List + +# --------------------------------------------------------------------------- +# VictoriaMetrics 连接 +# --------------------------------------------------------------------------- + +#: VM HTTP 地址，默认本机 +VM_URL: str = os.environ.get("PROTOFORGE_VM_URL", "http://localhost:8428") + +# --------------------------------------------------------------------------- +# 状态持久化 +# --------------------------------------------------------------------------- + +#: 健康模板状态文件路径（JSON），重启后可恢复学习进度 +STATE_FILE: str = os.environ.get( + "PROTOFORGE_STATE_FILE", + "/tmp/protoforge_predictor_state_v14.json", +) + +# --------------------------------------------------------------------------- +# 轮询与预测时间窗口 +# --------------------------------------------------------------------------- + +#: 拉取历史数据的时间窗口（分钟） +HISTORY_MINUTES: int = int(os.environ.get("PROTOFORGE_HISTORY_MINUTES", "30")) + +#: 理论预测时间跨度（秒） +HORIZON_SECONDS: int = int(os.environ.get("PROTOFORGE_HORIZON_SECONDS", "120")) + +#: 轮询间隔（秒） +POLL_INTERVAL: int = int(os.environ.get("PROTOFORGE_POLL_INTERVAL", "30")) + +#: 实际写入 VM 的预测点数 = min(HORIZON_SECONDS, POLL_INTERVAL) +#: 避免写入过多未来点导致 Grafana 图表出现"预测跳跃" +WRITE_HORIZON_SECONDS: int = min(HORIZON_SECONDS, POLL_INTERVAL) + +#: VM 查询步长 +QUERY_STEP: str = "1s" + +#: 最少需要多少个历史点才能开始建模 +MIN_POINTS: int = 120 + +# --------------------------------------------------------------------------- +# 周期检测范围 +# --------------------------------------------------------------------------- + +#: 允许的最短周期（秒） +MIN_PERIOD_SECONDS: int = 5 + +#: 允许的最长周期（秒） +MAX_PERIOD_SECONDS: int = 3600 + +# --------------------------------------------------------------------------- +# 模板学习参数 +# --------------------------------------------------------------------------- + +#: 构建模板至少需要多少个完整周期 +MIN_FULL_CYCLES_FOR_TEMPLATE: int = 3 + +#: 最多使用最近多少个周期来构建模板（防止过旧数据污染） +MAX_CYCLES_FOR_TEMPLATE: int = 8 + +#: 谷底检测时，低于此百分位的点才被视为谷底候选 +VALLEY_QUANTILE: int = 45 + +#: 健康状态下模板 EMA 更新步长（越小越保守，变化越慢） +HEALTHY_EMA_ALPHA: float = 0.10 + +#: 恢复状态下模板 EMA 更新步长（比健康状态更激进，加速追赶） +RECOVERY_EMA_ALPHA: float = 0.25 + +# --------------------------------------------------------------------------- +# 异常检测默认阈值 +# --------------------------------------------------------------------------- + +#: 检测窗口（秒）：只看最近这段时间的数据来判断是否异常 +DETECT_WINDOW_SECONDS: int = 30 + +#: 恢复期最短持续时间（秒）：异常消失后至少稳定这么久才恢复学习 +RECOVERY_MIN_SECONDS: int = 60 + +#: 越界比例阈值：窗口内超过此比例的点越界则报警 +OUTSIDE_RATIO_THRESHOLD: float = 0.60 + +#: 连续越界阈值（秒）：连续越界超过此秒数则报警 +MIN_CONSECUTIVE_OUTSIDE: int = 5 + +#: 严重越界倍数：单点超出边界宽度的此倍数则立即报警 +SEVERE_EXCEED_RATIO: float = 1.8 + +#: 真实数据最大允许延迟（秒）：超过此值认为数据管道异常，跳过预测 +MAX_DATA_LAG_SECONDS: int = 180 + +# --------------------------------------------------------------------------- +# Phase-lock 搜索参数 +# --------------------------------------------------------------------------- + +#: phase-lock 使用的最短历史窗口（秒） +PHASE_LOCK_MIN_WINDOW_SECONDS: int = 45 + +#: phase-lock 使用的最长历史窗口（秒） +PHASE_LOCK_MAX_WINDOW_SECONDS: int = 180 + +#: 周期搜索范围（相对于基准周期的比例），由 infer_metric_profile 动态覆盖 +PHASE_LOCK_PERIOD_SEARCH_RATIO: float = 0.12 + +#: 相位原点搜索范围（相对于周期的比例） +PHASE_LOCK_ORIGIN_SEARCH_RATIO: float = 0.35 + +#: 周期搜索步长（秒） +PHASE_LOCK_PERIOD_STEP: int = 1 + +#: 相位原点搜索步长（秒） +PHASE_LOCK_ORIGIN_STEP: int = 1 + +# --------------------------------------------------------------------------- +# 监控指标白名单 +# --------------------------------------------------------------------------- + +#: 默认监控的指标名列表 +_DEFAULT_MONITORED_METRICS: List[str] = [ + "feed_rate", + "spindle_speed", + "spindle_current", + "spindle_load", + "vibration_x", + "vibration_y", + "vibration_z", +] + +#: 实际使用的监控指标列表，可通过环境变量 PROTOFORGE_MONITORED_METRICS 覆盖 +#: 格式：逗号分隔的指标名，例如 "feed_rate,spindle_speed" +MONITORED_METRICS: List[str] = [ + m.strip() + for m in os.environ.get( + "PROTOFORGE_MONITORED_METRICS", + ",".join(_DEFAULT_MONITORED_METRICS), + ).split(",") + if m.strip() +] + +# --------------------------------------------------------------------------- +# 人工上下限覆盖文件 +# --------------------------------------------------------------------------- + +#: 覆盖文件路径，文件不存在时忽略（不报错） +#: 文件格式（JSON）： +#: { +#: "device-id": { +#: "metric_name": {"hard_max": 35.0, "hard_min": 0.0} +#: } +#: } +OVERRIDE_FILE: str = os.environ.get( + "PROTOFORGE_PREDICTOR_OVERRIDE", + "/etc/protoforge/predictor_override.json", +) + +# --------------------------------------------------------------------------- +# 目标列表刷新间隔 +# --------------------------------------------------------------------------- + +#: 每隔多少秒重新发现设备和指标（秒） +TARGETS_REFRESH_INTERVAL: int = int( + os.environ.get("PROTOFORGE_TARGETS_REFRESH", "60") +) + +# --------------------------------------------------------------------------- +# 写入 VM 时附加的额外标签 +# --------------------------------------------------------------------------- + +#: 附加到所有预测指标上的标签，用于在 Grafana 中区分预测数据和原始数据 +EXTRA_PREDICT_LABELS = { + "forecast": "phase_band_health_v14", + "source": "protoforge", +} diff --git a/ai/predictor/discovery.py b/ai/predictor/discovery.py new file mode 100644 index 0000000..c8b829a --- /dev/null +++ b/ai/predictor/discovery.py @@ -0,0 +1,84 @@ +# -*- coding: utf-8 -*- +""" +predictor.discovery +~~~~~~~~~~~~~~~~~~~ +Layer 1：设备与指标发现。 + +职责： +- 从 VictoriaMetrics 查询所有在线设备（device_id 标签值） +- 探测指定设备上哪些指标名当前有数据 + +本模块只做网络查询，不包含任何预测或统计逻辑。 + +依赖：requests +""" + +import logging +from typing import List + +import requests + +logger = logging.getLogger(__name__) + + +def discover_device_ids(vm_url: str) -> List[str]: + """ + 从 VictoriaMetrics 查询所有 device_id 标签值。 + + 调用 VM 的 label values 接口，返回当前存储中出现过的所有设备 ID。 + 网络失败时返回空列表，不抛出异常，由调用方决定如何处理。 + + Args: + vm_url: VM HTTP 地址，如 "http://localhost:8428" + + Returns: + 设备 ID 字符串列表，空字符串已过滤。 + 查询失败时返回空列表。 + """ + try: + resp = requests.get( + f"{vm_url}/api/v1/label/device_id/values", + timeout=10, + ) + resp.raise_for_status() + return [v for v in resp.json().get("data", []) if v] + except requests.RequestException as e: + logger.error("发现 device_id 失败: %s", e) + return [] + + +def discover_metrics_for_device( + vm_url: str, + device_id: str, + candidates: List[str], +) -> List[str]: + """ + 探测指定设备在 VictoriaMetrics 中实际存在且有近期数据的指标名。 + + 对 candidates 中的每个指标名发起即时查询（instant query）， + 只有返回非空 result 的指标才被认为"存在"。 + + Args: + vm_url: VM HTTP 地址 + device_id: 设备标识，对应 VM 中的 device_id 标签值 + candidates: 待探测的指标名列表，如 ["feed_rate", "spindle_speed"] + + Returns: + 实际有数据的指标名列表（保持 candidates 中的顺序）。 + 单个指标查询失败时静默跳过，不影响其他指标的探测。 + """ + found: List[str] = [] + for metric in candidates: + try: + resp = requests.get( + f"{vm_url}/api/v1/query", + params={"query": f'{metric}{{device_id="{device_id}"}}'}, + timeout=5, + ) + resp.raise_for_status() + if resp.json().get("data", {}).get("result"): + found.append(metric) + except requests.RequestException: + # 单个指标查询失败不影响整体发现流程 + pass + return found diff --git a/ai/predictor/models.py b/ai/predictor/models.py new file mode 100644 index 0000000..8b48bf5 --- /dev/null +++ b/ai/predictor/models.py @@ -0,0 +1,114 @@ +# -*- coding: utf-8 -*- +""" +predictor.models +~~~~~~~~~~~~~~~~ +纯数据结构定义，不包含任何业务逻辑或 IO 操作。 + +包含： +- ``BaselineState``：单个指标的健康模板状态，记录周期、模板曲线、健康状态等 +- ``MetricProfile``：从历史数据统计出的指标特征，驱动策略和阈值的自动推断 +- 状态常量：HEALTHY / ANOMALY / RECOVERING +""" + +from dataclasses import dataclass, field +from typing import List + +# --------------------------------------------------------------------------- +# 基线状态常量 +# --------------------------------------------------------------------------- + +#: 正常运行，模板持续学习更新 +BASELINE_STATUS_HEALTHY = "healthy" + +#: 检测到异常，模板冻结，不学习故障数据 +BASELINE_STATUS_ANOMALY = "anomaly" + +#: 异常消失，等待稳定后恢复学习 +BASELINE_STATUS_RECOVERING = "recovering" + + +# --------------------------------------------------------------------------- +# 基线状态 +# --------------------------------------------------------------------------- + +@dataclass +class BaselineState: + """ + 单个指标的健康基线状态。 + + 每个 (device_id, metric) 对应一个独立的 BaselineState 实例， + 存储该指标的周期模板和当前健康状态。 + + Attributes: + period: 检测到的加工周期长度（秒） + phase_origin_ts: 相位原点时间戳（Unix 秒），用于计算当前相位 + template: 中位数模板曲线，长度等于 period，用于预测 + lower_template: 下界模板曲线（phase_band 策略时为分位数，否则等于 template） + upper_template: 上界模板曲线（phase_band 策略时为分位数，否则等于 template） + strategy: 预测策略，"phase_point" 或 "phase_band" + status: 当前健康状态，取值为 BASELINE_STATUS_* 常量 + clean_seconds: 连续健康运行的秒数，用于判断是否可以更新模板 + last_update_ts: 上次模板更新的时间戳（Unix 秒） + last_seen_ts: 上次处理该指标的时间戳（Unix 秒），用于计算 elapsed + y_min: 最近一段时间内的最小值，用于量程参考 + y_max: 最近一段时间内的最大值，用于量程参考 + """ + + period: int + phase_origin_ts: int + template: List[float] + lower_template: List[float] + upper_template: List[float] + strategy: str + status: str + clean_seconds: int + last_update_ts: int + last_seen_ts: int + y_min: float + y_max: float + + +# --------------------------------------------------------------------------- +# 指标特征（自适应配置推断结果） +# --------------------------------------------------------------------------- + +@dataclass +class MetricProfile: + """ + 从历史数据统计出的指标特征，用于自动推断预测策略和阈值。 + + 由 ``profiling.infer_metric_profile()`` 生成， + 再由 ``profiling.build_target()`` 转换为执行层 target dict。 + + Attributes: + device_id: 设备标识，对应 VM 中的 device_id 标签值 + metric: 指标名，如 "feed_rate"、"spindle_current" + p5: 活跃段第 5 百分位数（过滤空闲零值后） + p95: 活跃段第 95 百分位数 + iqr: p95 - p5，反映正常波动范围 + cv: 变异系数（std / mean），衡量信号稳定性 + cv < 0.15 → 稳定信号（精铣类）→ phase_point + cv >= 0.15 → 波动信号（粗铣负载、振动类）→ phase_band + strategy: 自动推断的预测策略，"phase_point" 或 "phase_band" + abs_threshold: 绝对误差阈值（自动计算） + rel_threshold: 相对误差阈值（自动计算） + band_low_q: phase_band 下界分位数（默认 5） + band_high_q: phase_band 上界分位数（默认 95） + band_pad_abs: phase_band 额外填充宽度，覆盖正常尖峰 + phase_lock_period_search_ratio: phase-lock 周期搜索范围（相对比例） + 由实测周期抖动率动态决定，周期越不稳定则搜索范围越宽 + """ + + device_id: str + metric: str + p5: float + p95: float + iqr: float + cv: float + strategy: str + abs_threshold: float + rel_threshold: float + band_low_q: float + band_high_q: float + band_pad_abs: float + phase_lock_period_search_ratio: float diff --git a/ai/predictor/phase_lock.py b/ai/predictor/phase_lock.py new file mode 100644 index 0000000..b38acaf --- /dev/null +++ b/ai/predictor/phase_lock.py @@ -0,0 +1,152 @@ +# -*- coding: utf-8 -*- +""" +predictor.phase_lock +~~~~~~~~~~~~~~~~~~~~ +Phase-lock 相位对齐：在每次预测前动态校正周期和相位原点。 + +职责： +- 在基准周期附近搜索最优 (period, origin) 组合 +- 最小化最近时间窗口内的预测 MAE +- 支持 target 级别的搜索范围配置（粗铣工位周期抖动大，需要更宽的范围） + +依赖：predictor.template, predictor.config, predictor.models +""" + +import logging +from typing import Dict, Optional, Tuple + +import numpy as np + +from . import config +from .models import BaselineState +from .template import ( + normalize_origin_near, + predict_template_values, + resample_template, +) + +logger = logging.getLogger(__name__) + + +def phase_lock_recent( + state: BaselineState, + ts_grid: np.ndarray, + ys_model: np.ndarray, + target: Optional[Dict] = None, +) -> Tuple[int, int, np.ndarray, float]: + """ + 在最近时间窗口内搜索最优 (period, phase_origin) 组合。 + + 搜索策略： + 1. 确定搜索窗口（min/max 之间，约 2 倍周期） + 2. 在 [base_period * (1 - ratio), base_period * (1 + ratio)] 范围内枚举周期 + 3. 对每个周期，在 origin ± origin_shift 范围内枚举相位原点 + 4. 选择 MAE + 周期偏移惩罚最小的组合 + （惩罚项防止无谓地漂移到远离基准的周期） + + Args: + state: 当前基线状态（提供基准 period、origin、template） + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_model: 平滑后的信号（用于计算 MAE） + target: target dict，可包含 phase_lock_period_search_ratio / + phase_lock_origin_search_ratio 覆盖默认搜索范围 + + Returns: + (best_period, best_origin, best_pred, best_mae) 元组： + - best_period: 最优周期（整数秒） + - best_origin: 最优相位原点（Unix 秒） + - best_pred: 最优参数下的预测值数组（长度为搜索窗口大小） + - best_mae: 最优 MAE + """ + base_period = int(state.period) + base_origin = int(state.phase_origin_ts) + base_template = np.array(state.template, dtype=float) + + # 从 target 读取搜索范围，允许粗铣工位使用更宽的范围 + period_search_ratio = float( + (target or {}).get("phase_lock_period_search_ratio", config.PHASE_LOCK_PERIOD_SEARCH_RATIO) + ) + origin_search_ratio = float( + (target or {}).get("phase_lock_origin_search_ratio", config.PHASE_LOCK_ORIGIN_SEARCH_RATIO) + ) + + # 数据不足时直接返回基准预测 + if base_period <= 1 or len(base_template) <= 1: + ts_recent = ts_grid[-config.DETECT_WINDOW_SECONDS :].astype(int).tolist() + pred = predict_template_values(base_template, base_period, base_origin, ts_recent) + actual = ys_model[-len(ts_recent) :].astype(float) + mae = float(np.mean(np.abs(actual - pred))) if len(actual) else 0.0 + return base_period, base_origin, pred, mae + + # 搜索窗口：约 2 倍周期，clip 到 [min, max] + window_seconds = max( + config.PHASE_LOCK_MIN_WINDOW_SECONDS, + min(config.PHASE_LOCK_MAX_WINDOW_SECONDS, int(base_period * 2)), + ) + + cutoff = ts_grid[-1] - window_seconds + mask = ts_grid >= cutoff + ts_recent_arr = ts_grid[mask].astype(int) + actual = ys_model[mask].astype(float) + + # 窗口内数据不足时退化到固定长度 + if len(ts_recent_arr) < max(10, config.DETECT_WINDOW_SECONDS): + ts_recent_arr = ts_grid[-config.DETECT_WINDOW_SECONDS :].astype(int) + actual = ys_model[-config.DETECT_WINDOW_SECONDS :].astype(float) + + ts_recent = ts_recent_arr.tolist() + last_ts = int(ts_recent[-1]) + + # 周期搜索范围 + p_min = max( + int(config.MIN_PERIOD_SECONDS), + int(round(base_period * (1.0 - period_search_ratio))), + ) + p_max = min( + int(config.MAX_PERIOD_SECONDS), + int(round(base_period * (1.0 + period_search_ratio))), + ) + + # 初始化为基准参数 + best_period = base_period + best_origin = normalize_origin_near(base_origin, base_period, last_ts) + best_template = resample_template(base_template, best_period) + best_pred = predict_template_values( + template=best_template, + period=best_period, + phase_origin_ts=best_origin, + ts_list=ts_recent, + ) + best_mae = float(np.mean(np.abs(actual - best_pred))) + + for period in range(p_min, p_max + 1, config.PHASE_LOCK_PERIOD_STEP): + template = resample_template(base_template, period) + center_origin = normalize_origin_near(base_origin, period, last_ts) + origin_shift = max(2, int(round(period * origin_search_ratio))) + + for shift in range(-origin_shift, origin_shift + 1, config.PHASE_LOCK_ORIGIN_STEP): + origin = center_origin + shift + pred = predict_template_values( + template=template, + period=period, + phase_origin_ts=origin, + ts_list=ts_recent, + ) + mae = float(np.mean(np.abs(actual - pred))) + + # 惩罚项：偏离基准周期越远，惩罚越大（0.5 秒/秒偏差） + # 防止在噪声中漂移到远离真实周期的位置 + penalty = abs(period - base_period) * 0.5 + score = mae + penalty + best_score = best_mae + abs(best_period - base_period) * 0.5 + + if score < best_score: + best_period = period + best_origin = origin + best_pred = pred + best_mae = mae + + # 规整化最终原点到最新时间戳附近 + best_origin = normalize_origin_near(best_origin, best_period, last_ts) + + return int(best_period), int(best_origin), best_pred, float(best_mae) diff --git a/ai/predictor/profiling.py b/ai/predictor/profiling.py new file mode 100644 index 0000000..5330500 --- /dev/null +++ b/ai/predictor/profiling.py @@ -0,0 +1,256 @@ +# -*- coding: utf-8 -*- +""" +predictor.profiling +~~~~~~~~~~~~~~~~~~~ +Layer 2：自适应配置推断。 + +职责： +- 从历史数据统计指标特征（p5/p95/IQR/cv/周期抖动率） +- 自动推断预测策略（phase_point vs phase_band）和阈值 +- 加载人工上下限覆盖文件（override） +- 将 MetricProfile 转换为执行层 target dict +- 完整的发现 + 推断流程（refresh_targets） + +依赖：predictor.storage, predictor.discovery, predictor.signal, predictor.models, predictor.config +""" + +import json +import logging +import os +from typing import Dict, List, Optional + +import numpy as np + +from . import config +from .discovery import discover_device_ids, discover_metrics_for_device +from .models import MetricProfile +from .signal import estimate_period_rough, find_valley_indices +from .storage import fetch_history, normalize_history + +logger = logging.getLogger(__name__) + + +def infer_metric_profile( + vm_url: str, + device_id: str, + metric: str, +) -> Optional[MetricProfile]: + """ + 拉取历史数据，统计活跃段特征，自动推断预测策略和阈值。 + + 推断逻辑： + - 空闲段过滤：排除 p10 以下的点，避免机床空闲时的零值拉低阈值 + - strategy 判断：cv < 0.15 → phase_point（稳定信号），否则 phase_band（波动信号） + - abs_threshold：取 IQR * 0.8、量程 * 0.05、std * 2.0 三者最大值 + - rel_threshold：min(0.30, cv * 1.5) + - band_pad_abs：max(IQR * 0.3, std)，覆盖正常尖峰 + - phase_lock_period_search_ratio：由周期抖动率动态决定，clip 到 [0.12, 0.25] + + Args: + vm_url: VM HTTP 地址 + device_id: 设备标识 + metric: 指标名 + + Returns: + MetricProfile，数据不足时返回 None。 + """ + ts_raw, ys_raw = fetch_history( + vm_url=vm_url, + query=f'{metric}{{device_id="{device_id}"}}', + minutes=config.HISTORY_MINUTES, + ) + + if len(ys_raw) < config.MIN_POINTS: + return None + + arr = np.array(ys_raw, dtype=float) + + # 过滤空闲段：只保留活跃值（高于 p10） + p10_val = float(np.percentile(arr, 10)) + active = arr[arr > p10_val] + if len(active) < 30: + active = arr # 数据全是活跃段，不过滤 + + mean_val = float(np.mean(active)) + std_val = float(np.std(active)) + cv = std_val / max(abs(mean_val), 1e-6) + p5 = float(np.percentile(active, 5)) + p95 = float(np.percentile(active, 95)) + iqr = p95 - p5 + + # 策略自动判断：cv 衡量信号稳定性 + strategy = "phase_point" if cv < 0.15 else "phase_band" + + # 阈值自动计算 + abs_threshold = max(iqr * 0.8, (p95 - p5) * 0.05, std_val * 2.0) + rel_threshold = min(0.30, cv * 1.5) + + # phase_band 容忍带宽度 + band_pad_abs = max(iqr * 0.3, std_val) + + # phase-lock 搜索范围：从历史数据估算周期抖动率 + ts_grid, ys_grid = normalize_history(ts_raw, ys_raw) + period_search_ratio = config.PHASE_LOCK_PERIOD_SEARCH_RATIO # 默认值 + + if len(ys_grid) >= config.MIN_POINTS: + rough_period = estimate_period_rough(ys_grid) + if rough_period > config.MIN_PERIOD_SECONDS: + valleys = find_valley_indices(ts_grid, ys_grid, rough_period) + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys].astype(float)) + valid = diffs[ + (diffs > rough_period * 0.5) & (diffs < rough_period * 2.0) + ] + if len(valid) >= 2: + # 周期变异系数 * 2 作为搜索范围，clip 到 [0.12, 0.25] + period_cv = float(np.std(valid) / max(np.mean(valid), 1e-6)) + period_search_ratio = float(np.clip(period_cv * 2.0, 0.12, 0.25)) + + logger.info( + "推断指标特征 device=%s metric=%s cv=%.3f strategy=%s " + "abs_thr=%.3f rel_thr=%.3f period_search=%.2f", + device_id, metric, cv, strategy, + abs_threshold, rel_threshold, period_search_ratio, + ) + + return MetricProfile( + device_id=device_id, + metric=metric, + p5=p5, + p95=p95, + iqr=iqr, + cv=cv, + strategy=strategy, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + band_low_q=5.0, + band_high_q=95.0, + band_pad_abs=band_pad_abs, + phase_lock_period_search_ratio=period_search_ratio, + ) + + +def load_overrides(path: str) -> Dict: + """ + 加载人工上下限覆盖文件，文件不存在时返回空字典。 + + 文件格式（JSON）： + { + "device-id": { + "metric_name": {"hard_max": 35.0, "hard_min": 0.0} + } + } + + Args: + path: 覆盖文件路径 + + Returns: + 覆盖配置字典，文件不存在或解析失败时返回空字典。 + """ + if not os.path.exists(path): + return {} + try: + with open(path, "r", encoding="utf-8") as f: + return json.load(f) + except Exception as e: + logger.warning("加载 override 文件失败 %s: %s", path, e) + return {} + + +def build_target(profile: MetricProfile, overrides: Dict) -> Dict: + """ + 将 MetricProfile 转换为预测执行层可用的 target dict。 + + target dict 包含 run_once() 所需的全部配置： + - query / pred_metric / anomaly_metric + - strategy / 阈值 / 平滑窗口 + - phase-lock 搜索范围 + - 物理上下限（可选，来自 override 文件） + + Args: + profile: 从历史数据推断出的指标特征 + overrides: 覆盖配置字典（来自 load_overrides） + + Returns: + target dict。 + """ + device_overrides = overrides.get(profile.device_id, {}).get(profile.metric, {}) + + target: Dict = { + "query": f'{profile.metric}{{device_id="{profile.device_id}"}}', + "pred_metric": f"{profile.metric}_predicted", + "anomaly_metric": f"{profile.metric}_anomaly", + "strategy": profile.strategy, + "abs_threshold": profile.abs_threshold, + "rel_threshold": profile.rel_threshold, + # phase_band 用更大的平滑窗口抑制脉冲噪声 + "smooth_window": 5 if profile.strategy == "phase_band" else 2, + "outside_ratio_threshold": 0.60, + "min_consecutive_outside": 5, + "severe_exceed_ratio": 1.8, + "phase_lock_period_search_ratio": profile.phase_lock_period_search_ratio, + # origin 搜索范围约为 period 搜索范围的 2.5 倍 + "phase_lock_origin_search_ratio": min( + 0.45, profile.phase_lock_period_search_ratio * 2.5 + ), + # 物理上下限（可选，来自 override 文件） + "hard_max": device_overrides.get("hard_max"), + "hard_min": device_overrides.get("hard_min"), + } + + if profile.strategy == "phase_band": + target.update({ + "band_low_q": profile.band_low_q, + "band_high_q": profile.band_high_q, + "band_pad_abs": profile.band_pad_abs, + }) + + return target + + +def refresh_targets( + vm_url: str, + monitored_metrics: List[str], + override_path: str, +) -> List[Dict]: + """ + 完整的发现 + 推断流程：发现所有设备，推断所有指标的配置，返回 target list。 + + 流程： + 1. 从 VM 发现所有 device_id + 2. 对每个设备探测哪些指标有数据 + 3. 对每个有数据的指标推断 MetricProfile + 4. 将 MetricProfile 转换为 target dict + + 若发现失败（无 device_id），返回空列表，由调用方决定是否保留旧列表。 + + Args: + vm_url: VM HTTP 地址 + monitored_metrics: 待探测的指标名候选列表 + override_path: 覆盖文件路径 + + Returns: + target dict 列表，每个元素对应一个 (device_id, metric) 对。 + """ + logger.info("开始发现设备和指标...") + overrides = load_overrides(override_path) + targets: List[Dict] = [] + + device_ids = discover_device_ids(vm_url) + if not device_ids: + logger.warning("未发现任何 device_id") + return [] + + for device_id in device_ids: + metrics = discover_metrics_for_device(vm_url, device_id, monitored_metrics) + for metric in metrics: + profile = infer_metric_profile(vm_url, device_id, metric) + if profile is not None: + targets.append(build_target(profile, overrides)) + + logger.info( + "目标列表已更新：%d 台设备，%d 个指标目标", + len(device_ids), + len(targets), + ) + return targets diff --git a/ai/predictor/service.py b/ai/predictor/service.py new file mode 100644 index 0000000..1e9dca7 --- /dev/null +++ b/ai/predictor/service.py @@ -0,0 +1,333 @@ +# -*- coding: utf-8 -*- +""" +predictor.service +~~~~~~~~~~~~~~~~~ +主服务类：组装所有模块，驱动预测主循环。 + +职责： +- 持有所有运行时状态（baseline_states、last_written、targets） +- 按 TARGETS_REFRESH_INTERVAL 定期重新发现设备和指标 +- 每轮轮询：拉取历史数据 → 更新状态 → 预测 → 写入 VM +- 每轮结束后持久化状态到文件 + +依赖：所有其他 predictor 子模块 +""" + +import logging +import time +from datetime import datetime +from typing import Dict, List, Optional + +from . import config +from .anomaly import calc_final_bounds +from .models import BaselineState +from .profiling import refresh_targets +from .signal import preprocess_values +from .state import maybe_update_state +from .storage import ( + fetch_history, + load_state, + merge_labels, + normalize_history, + parse_labels_from_query, + save_state, + series_key, + write_prediction_bundle, +) +from .template import predict_state_bundle + +logger = logging.getLogger(__name__) + + +class PredictorService: + """ + 预测服务主类。 + + 封装所有运行时状态，支持多实例部署（每个实例独立持有状态）。 + 通过 run() 启动主循环，通过 run_once() 执行单轮预测。 + + Attributes: + _vm_url: VM HTTP 地址 + _state_file: 状态持久化文件路径 + _history_minutes: 拉取历史数据的时间窗口（分钟） + _write_horizon: 实际写入 VM 的预测点数（秒） + _poll_interval: 轮询间隔（秒） + _targets_refresh_interval: 目标列表刷新间隔（秒） + _monitored_metrics: 待监控的指标名列表 + _override_file: 人工上下限覆盖文件路径 + _extra_labels: 写入 VM 时附加的额外标签 + _states: key → BaselineState 的字典（运行时状态） + _last_written: key → 上次写入的真实数据时间戳 + _targets: 当前目标列表 + _targets_last_refresh: 上次刷新目标列表的时间戳 + """ + + def __init__( + self, + vm_url: str = config.VM_URL, + state_file: str = config.STATE_FILE, + history_minutes: int = config.HISTORY_MINUTES, + write_horizon: int = config.WRITE_HORIZON_SECONDS, + poll_interval: int = config.POLL_INTERVAL, + targets_refresh_interval: int = config.TARGETS_REFRESH_INTERVAL, + monitored_metrics: Optional[List[str]] = None, + override_file: str = config.OVERRIDE_FILE, + extra_labels: Optional[Dict[str, str]] = None, + ) -> None: + self._vm_url = vm_url + self._state_file = state_file + self._history_minutes = history_minutes + self._write_horizon = write_horizon + self._poll_interval = poll_interval + self._targets_refresh_interval = targets_refresh_interval + self._monitored_metrics = monitored_metrics or config.MONITORED_METRICS + self._override_file = override_file + self._extra_labels = extra_labels or config.EXTRA_PREDICT_LABELS + + self._states: Dict[str, BaselineState] = {} + self._last_written: Dict[str, int] = {} + self._targets: List[Dict] = [] + self._targets_last_refresh: float = 0.0 + + # ------------------------------------------------------------------ + # 目标列表管理 + # ------------------------------------------------------------------ + + def _refresh_targets_if_needed(self) -> None: + """ + 按 targets_refresh_interval 间隔重新发现设备和指标。 + + 首次调用时立即执行发现。发现失败时保留现有目标列表。 + """ + now = time.time() + if now - self._targets_last_refresh < self._targets_refresh_interval and self._targets: + return + + new_targets = refresh_targets( + vm_url=self._vm_url, + monitored_metrics=self._monitored_metrics, + override_path=self._override_file, + ) + + if new_targets: + self._targets = new_targets + self._targets_last_refresh = now + else: + logger.warning("发现流程未产生任何有效目标，保持现有目标列表") + + # ------------------------------------------------------------------ + # 预测时间轴 + # ------------------------------------------------------------------ + + def _build_prediction_timestamps( + self, + key: str, + last_real_ts: int, + now_sec: int, + ) -> Optional[List[int]]: + """ + 构建预测时间戳列表（从 last_real_ts + 1 开始，共 write_horizon 个点）。 + + 两种情况下跳过写入： + 1. 真实数据延迟过大（数据管道异常） + 2. 真实数据时间戳未推进（重复写入同一批预测） + + Args: + key: 序列标识符 + last_real_ts: 最新真实数据点的时间戳（Unix 秒） + now_sec: 当前时间戳（Unix 秒） + + Returns: + 预测时间戳列表，跳过时返回 None。 + """ + data_lag = now_sec - last_real_ts + + if data_lag > config.MAX_DATA_LAG_SECONDS: + logger.warning( + "真实数据延迟过大，跳过预测 key=%s data_lag=%ss max=%ss", + key, data_lag, config.MAX_DATA_LAG_SECONDS, + ) + return None + + last_written_real_ts = self._last_written.get(key) + if last_written_real_ts is not None and last_real_ts <= int(last_written_real_ts): + logger.info( + "真实数据时间戳未推进，跳过重复写入 key=%s last_real_ts=%s last_written=%s", + key, last_real_ts, last_written_real_ts, + ) + return None + + return [last_real_ts + i + 1 for i in range(self._write_horizon)] + + # ------------------------------------------------------------------ + # 单轮预测 + # ------------------------------------------------------------------ + + def run_once(self) -> None: + """ + 执行一轮预测：遍历所有目标，拉取数据、更新状态、写入预测结果。 + + 每轮结束后将状态持久化到文件。 + """ + now_str = datetime.now().strftime("%H:%M:%S") + + self._refresh_targets_if_needed() + + if not self._targets: + logger.warning("[%s] 目标列表为空，等待设备发现完成", now_str) + return + + for target in self._targets: + query = target["query"] + pred_metric = target["pred_metric"] + anomaly_metric = target["anomaly_metric"] + strategy = target.get("strategy", "phase_point") + smooth_window = int(target.get("smooth_window", 1)) + + # 1. 拉取历史数据 + ts, ys = fetch_history( + vm_url=self._vm_url, + query=query, + minutes=self._history_minutes, + ) + + if len(ys) < config.MIN_POINTS: + logger.info("[%s] %s 数据不足（%d 点），跳过", now_str, query, len(ys)) + continue + + ts_grid, ys_grid_raw = normalize_history(ts, ys) + + if len(ys_grid_raw) < config.MIN_POINTS: + logger.info( + "[%s] %s 清洗后数据不足（%d 点），跳过", + now_str, query, len(ys_grid_raw), + ) + continue + + # 2. 预处理（平滑） + ys_grid_model = preprocess_values(ys_grid_raw, strategy, smooth_window) + + # 3. 构建写入标签 + base_labels = parse_labels_from_query(query) + write_labels = merge_labels(base_labels, self._extra_labels) + key = series_key(pred_metric, write_labels) + + # 4. 更新状态（异常检测 + 模板学习） + ( + state, + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + max_outside_seconds, + max_exceed_ratio, + ) = maybe_update_state( + key=key, + ts_grid=ts_grid, + ys_model=ys_grid_model, + ys_actual=ys_grid_raw, + target=target, + states=self._states, + ) + + if state is None: + logger.info("[%s] %s 暂无可用健康模板，等待学习", now_str, query) + continue + + # 5. 构建预测时间戳 + now_sec = int(time.time()) + last_real_ts = int(ts_grid[-1]) + data_lag = now_sec - last_real_ts + + ts_future = self._build_prediction_timestamps( + key=key, + last_real_ts=last_real_ts, + now_sec=now_sec, + ) + + if not ts_future: + continue + + # 6. 预测 + pred_values, lower_raw, upper_raw = predict_state_bundle(state, ts_future) + lower_values, upper_values = calc_final_bounds( + state=state, + pred=pred_values, + lower_raw=lower_raw, + upper_raw=upper_raw, + target=target, + ) + + # 7. 写入 VM + ok = write_prediction_bundle( + vm_url=self._vm_url, + pred_metric=pred_metric, + anomaly_metric=anomaly_metric, + labels=write_labels, + ts_future=ts_future, + pred_values=pred_values, + lower_values=lower_values, + upper_values=upper_values, + is_anomaly=is_anomaly, + outside_ratio=outside_ratio, + mean_abs_err=mean_abs_err, + mean_rel_err=mean_rel_err, + max_outside_seconds=max_outside_seconds, + max_exceed_ratio=max_exceed_ratio, + event_ts=last_real_ts, + ) + + if not ok: + logger.error("[%s] %s 写入预测数据失败", now_str, query) + continue + + self._last_written[key] = last_real_ts + + # 8. 打印摘要日志 + future_start = datetime.fromtimestamp(ts_future[0]).strftime("%H:%M:%S") + future_end = datetime.fromtimestamp(ts_future[-1]).strftime("%H:%M:%S") + last_real_str = datetime.fromtimestamp(last_real_ts).strftime("%H:%M:%S") + origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") + + logger.info( + "[%s] %-50s → %-35s strategy=%s status=%s anomaly=%s " + "outside=%.2f max_outside=%ss max_exceed=%.2f " + "period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", + now_str, query, pred_metric, + state.strategy, state.status, is_anomaly, + outside_ratio, max_outside_seconds, max_exceed_ratio, + state.period, origin_str, last_real_str, data_lag, + len(ts_future), future_start, future_end, + ) + + save_state(self._state_file, self._states) + + # ------------------------------------------------------------------ + # 主循环 + # ------------------------------------------------------------------ + + def run(self) -> None: + """ + 启动预测服务主循环。 + + 加载持久化状态后进入无限循环，每隔 poll_interval 秒执行一次 run_once()。 + """ + self._states = load_state(self._state_file) + + logger.info( + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds " + "轮询间隔=%ds state=%s forecast=%s override=%s refresh=%ds", + self._vm_url, + self._history_minutes, + config.HORIZON_SECONDS, + self._write_horizon, + self._poll_interval, + self._state_file, + self._extra_labels.get("forecast", ""), + self._override_file, + self._targets_refresh_interval, + ) + + while True: + self.run_once() + time.sleep(self._poll_interval) diff --git a/ai/predictor/signal.py b/ai/predictor/signal.py new file mode 100644 index 0000000..e948160 --- /dev/null +++ b/ai/predictor/signal.py @@ -0,0 +1,335 @@ +# -*- coding: utf-8 -*- +""" +predictor.signal +~~~~~~~~~~~~~~~~ +纯信号处理与周期估计，不包含任何 IO 操作。 + +职责： +- 滚动中位数、移动平均等平滑算法 +- 基于 FFT + 自相关的周期估计 +- 谷底检测（用于模板构建的相位对齐） +- 原始数据预处理（根据策略选择平滑方式） + +本模块所有函数均为纯函数，输入 numpy 数组，输出 numpy 数组或基本类型。 + +依赖：numpy +""" + +import math +from typing import Dict, List, Tuple + +import numpy as np + +from . import config + + +def rolling_median(arr: np.ndarray, window: int) -> np.ndarray: + """ + 对数组做滚动中位数平滑（边缘用 edge 填充）。 + + 中位数对脉冲噪声鲁棒，适合 phase_band 策略的粗铣负载信号。 + window 自动调整为奇数，保证对称填充。 + + Args: + arr: 输入数组 + window: 滑动窗口大小（秒），<=1 时直接返回原数组 + + Returns: + 平滑后的数组，长度与输入相同。 + """ + if window <= 1 or len(arr) < window: + return arr.astype(float) + + # 保证奇数窗口，使填充对称 + if window % 2 == 0: + window += 1 + + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + result = [float(np.median(padded[i : i + window])) for i in range(len(arr))] + return np.array(result, dtype=float) + + +def moving_average(arr: np.ndarray, window: int) -> np.ndarray: + """ + 对数组做均匀权重移动平均（边缘用 edge 填充）。 + + 比滚动中位数快，适合 phase_point 策略的稳定信号。 + window 自动调整为奇数，保证对称填充。 + + Args: + arr: 输入数组 + window: 滑动窗口大小（秒），<=1 时直接返回原数组 + + Returns: + 平滑后的数组，长度与输入相同。 + """ + if window <= 1 or len(arr) < window: + return arr.astype(float) + + if window % 2 == 0: + window += 1 + + kernel = np.ones(window, dtype=float) / window + pad = window // 2 + padded = np.pad(arr.astype(float), (pad, pad), mode="edge") + + return np.convolve(padded, kernel, mode="valid") + + +def preprocess_values( + ys_grid: np.ndarray, + strategy: str, + smooth_window: int, +) -> np.ndarray: + """ + 根据预测策略对原始数据进行预处理平滑。 + + - phase_band：使用滚动中位数（对脉冲噪声鲁棒） + - phase_point：使用移动平均（保留趋势，计算更快） + + Args: + ys_grid: 均匀 1 秒网格上的原始值数组 + strategy: "phase_point" 或 "phase_band" + smooth_window: 平滑窗口大小（秒），<=1 时不平滑 + + Returns: + 平滑后的数组，长度与输入相同。 + """ + if strategy == "phase_band": + return rolling_median(ys_grid, smooth_window) + + if smooth_window > 1: + return moving_average(ys_grid, smooth_window) + + return ys_grid.astype(float) + + +# --------------------------------------------------------------------------- +# 周期估计 +# --------------------------------------------------------------------------- + +def estimate_period_by_fft(ys_arr: np.ndarray) -> float: + """ + 用 FFT 粗估信号的主周期（秒）。 + + 取去均值后的功率谱中能量最大的频率分量，转换为周期。 + 结果被 clip 到 [MIN_PERIOD_SECONDS, MAX_PERIOD_SECONDS]。 + + Args: + ys_arr: 均匀采样的值数组（1 秒间隔） + + Returns: + 估计的周期（秒），浮点数。数据不足或全零时返回 60.0。 + """ + n = len(ys_arr) + if n < 8: + return 60.0 + + centered = ys_arr - np.mean(ys_arr) + if np.allclose(centered, 0): + return 60.0 + + fft_vals = np.fft.rfft(centered) + freqs = np.fft.rfftfreq(n, d=1.0) + + if len(freqs) <= 1: + return 60.0 + + # 跳过直流分量（index 0），找功率最大的频率 + power = np.abs(fft_vals[1:]) + if len(power) == 0 or np.max(power) <= 0: + return 60.0 + + dominant_idx = int(np.argmax(power)) + 1 + dominant_freq = float(freqs[dominant_idx]) + + if dominant_freq <= 0: + return 60.0 + + period = 1.0 / dominant_freq + return float(np.clip(period, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + +def refine_period_by_autocorr(ys_arr: np.ndarray, init_period: float) -> float: + """ + 用自相关函数在 FFT 粗估周期附近精化周期。 + + 在 [init_period * 0.7, init_period * 1.3] 范围内搜索自相关峰值， + 比 FFT 对非整数周期和噪声更鲁棒。 + + Args: + ys_arr: 均匀采样的值数组 + init_period: FFT 粗估的初始周期（秒） + + Returns: + 精化后的周期（秒），clip 到合法范围。 + """ + n = len(ys_arr) + if n < 20: + return float(np.clip(init_period, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + centered = ys_arr - np.mean(ys_arr) + if np.allclose(centered, 0): + return float(np.clip(init_period, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + # 全相关，取正半轴（lag >= 0） + corr = np.correlate(centered, centered, mode="full")[n - 1:] + + p0 = int(round(init_period)) + left = max(int(config.MIN_PERIOD_SECONDS), int(max(2, p0 * 0.7))) + right = min(n // 2, int(max(left + 1, p0 * 1.3))) + + if right <= left: + return float(np.clip(init_period, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + search = corr[left : right + 1] + if len(search) == 0: + return float(np.clip(init_period, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + best_lag = left + int(np.argmax(search)) + return float(np.clip(best_lag, config.MIN_PERIOD_SECONDS, config.MAX_PERIOD_SECONDS)) + + +def estimate_period_rough(ys_arr: np.ndarray) -> int: + """ + FFT + 自相关两步法估计信号周期，返回整数秒。 + + 先用 FFT 粗估，再用自相关精化，最后 clip 到合法范围。 + + Args: + ys_arr: 均匀采样的值数组 + + Returns: + 估计的周期（整数秒）。 + """ + p_fft = estimate_period_by_fft(ys_arr) + p_refined = refine_period_by_autocorr(ys_arr, p_fft) + period = int(round(p_refined)) + period = max(int(config.MIN_PERIOD_SECONDS), min(int(config.MAX_PERIOD_SECONDS), period)) + return int(period) + + +# --------------------------------------------------------------------------- +# 谷底检测 +# --------------------------------------------------------------------------- + +def find_valley_indices( + ts_grid: np.ndarray, + ys_grid: np.ndarray, + expected_period: int, +) -> List[int]: + """ + 在时序数据中检测周期性谷底（加工周期的起始点）。 + + 算法步骤： + 1. 对信号做轻度移动平均平滑，抑制高频噪声 + 2. 找低于 VALLEY_QUANTILE 百分位的局部极小值作为候选 + 3. 若候选不足，放宽条件（不限百分位） + 4. 按最小间距过滤，同一间距内保留最低点 + 5. 按周期合理性（0.55~1.60 倍期望周期）清洗 + + Args: + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_grid: 对应的值数组 + expected_period: 预期周期（秒），用于设置最小间距和合理性检查 + + Returns: + 谷底在数组中的索引列表（已排序）。 + 数据不足时返回空列表。 + """ + n = len(ys_grid) + if n < max(10, expected_period * 2): + return [] + + period = max(3, int(expected_period)) + # 平滑窗口约为周期的 8%，最大 21 秒，避免过度平滑 + smooth_window = min(max(3, int(round(period * 0.08))), 21) + ys_smooth = moving_average(ys_grid, smooth_window) + + threshold = float(np.percentile(ys_smooth, config.VALLEY_QUANTILE)) + + # 第一轮：只取低于阈值的局部极小值 + candidates = [ + i for i in range(1, n - 1) + if ( + ys_smooth[i] <= ys_smooth[i - 1] + and ys_smooth[i] < ys_smooth[i + 1] + and ys_smooth[i] <= threshold + ) + ] + + # 候选不足时放宽：取所有局部极小值 + if len(candidates) < config.MIN_FULL_CYCLES_FOR_TEMPLATE: + candidates = [ + i for i in range(1, n - 1) + if ys_smooth[i] <= ys_smooth[i - 1] and ys_smooth[i] < ys_smooth[i + 1] + ] + + if not candidates: + return [] + + # 按最小间距过滤：同一间距内保留最低点 + min_distance = max(2, int(round(period * 0.55))) + selected: List[int] = [] + for idx in candidates: + if not selected: + selected.append(idx) + elif idx - selected[-1] >= min_distance: + selected.append(idx) + elif ys_smooth[idx] < ys_smooth[selected[-1]]: + selected[-1] = idx + + if len(selected) < 2: + return selected + + # 按周期合理性清洗：间距过小则保留更低点，间距过大则直接接受 + cleaned = [selected[0]] + for idx in selected[1:]: + diff = int(ts_grid[idx] - ts_grid[cleaned[-1]]) + if int(period * 0.55) <= diff <= int(period * 1.60): + cleaned.append(idx) + elif diff < int(period * 0.55): + # 间距太小，保留更低的那个 + if ys_smooth[idx] < ys_smooth[cleaned[-1]]: + cleaned[-1] = idx + else: + # 间距过大（可能漏检了一个谷底），直接接受 + cleaned.append(idx) + + return cleaned + + +def detect_period_and_valleys( + ts_grid: np.ndarray, + ys_grid: np.ndarray, +) -> Tuple[int, List[int]]: + """ + 综合估计周期并检测谷底。 + + 先粗估周期，再检测谷底，最后用谷底间距的中位数修正周期。 + 谷底间距的中位数比 FFT 更能反映实际加工节拍。 + + Args: + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_grid: 对应的值数组 + + Returns: + (period, valley_indices) 元组： + - period: 修正后的周期（整数秒） + - valley_indices: 谷底索引列表 + """ + rough = estimate_period_rough(ys_grid) + valleys = find_valley_indices(ts_grid, ys_grid, rough) + + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys]) + # 只取合理范围内的间距参与中位数计算 + good = diffs[(diffs >= rough * 0.55) & (diffs <= rough * 1.60)] + period = int(round(float(np.median(good)))) if len(good) > 0 else rough + else: + period = rough + + period = max(int(config.MIN_PERIOD_SECONDS), min(int(config.MAX_PERIOD_SECONDS), period)) + return int(period), valleys diff --git a/ai/predictor/state.py b/ai/predictor/state.py new file mode 100644 index 0000000..d7adedd --- /dev/null +++ b/ai/predictor/state.py @@ -0,0 +1,328 @@ +# -*- coding: utf-8 -*- +""" +predictor.state +~~~~~~~~~~~~~~~ +状态生命周期管理：BaselineState 的创建、更新和 phase-lock 应用。 + +职责： +- 首次见到某指标时初始化健康基线 +- 每轮轮询时运行异常检测，更新状态机（HEALTHY / ANOMALY / RECOVERING） +- 健康/恢复状态下用 EMA 渐进更新模板 +- 将 phase-lock 结果写回 state + +本模块不做任何 IO，states 字典由调用方（service.py）持有和传入。 + +依赖：predictor.template, predictor.anomaly, predictor.config, predictor.models +""" + +import logging +import time +from datetime import datetime +from typing import Dict, Optional, Tuple + +import numpy as np + +from . import config +from .anomaly import detect_anomaly +from .models import ( + BASELINE_STATUS_ANOMALY, + BASELINE_STATUS_HEALTHY, + BASELINE_STATUS_RECOVERING, + BaselineState, +) +from .template import ( + build_current_baseline, + merge_template, + resample_template, +) + +logger = logging.getLogger(__name__) + + +def create_initial_state( + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, + now_sec: int, +) -> Optional[BaselineState]: + """ + 从历史数据构建初始健康基线状态。 + + 首次见到某指标时调用，需要足够的历史数据（MIN_POINTS 个点）。 + + Args: + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_model: 平滑后的信号（用于周期检测和中值模板） + ys_actual: 原始信号（用于分位数模板和量程统计） + target: target dict，包含策略和阈值配置 + now_sec: 当前时间戳（Unix 秒） + + Returns: + 初始化的 BaselineState，数据不足时返回 None。 + """ + strategy = str(target.get("strategy", "phase_point")) + band_low_q = float(target.get("band_low_q", 5.0)) + band_high_q = float(target.get("band_high_q", 95.0)) + + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_mid_grid=ys_model, + ys_band_grid=ys_actual, + strategy=strategy, + band_low_q=band_low_q, + band_high_q=band_high_q, + ) + + if baseline is None: + return None + + period, phase_origin_ts, template, lower_template, upper_template = baseline + + return BaselineState( + period=int(period), + phase_origin_ts=int(phase_origin_ts), + template=template.astype(float).tolist(), + lower_template=lower_template.astype(float).tolist(), + upper_template=upper_template.astype(float).tolist(), + strategy=strategy, + status=BASELINE_STATUS_HEALTHY, + # 初始 clean_seconds 设为多个完整周期，表示已有足够的健康历史 + clean_seconds=int(period * config.MAX_CYCLES_FOR_TEMPLATE), + last_update_ts=now_sec, + last_seen_ts=now_sec, + y_min=float(np.min(ys_actual)), + y_max=float(np.max(ys_actual)), + ) + + +def apply_phase_lock_to_state( + state: BaselineState, + best_period: int, + best_origin: int, +) -> None: + """ + 将 phase-lock 搜索结果写回 state（原地修改）。 + + 若周期发生变化，同时对三条模板做重采样，保持长度一致。 + + Args: + state: 要更新的基线状态（原地修改） + best_period: phase-lock 找到的最优周期（整数秒） + best_origin: phase-lock 找到的最优相位原点（Unix 秒） + """ + best_period = int(best_period) + if best_period <= 1: + return + + # 周期变化时重采样三条模板 + if len(state.template) != best_period: + state.template = resample_template( + np.array(state.template, dtype=float), best_period + ).astype(float).tolist() + + if len(state.lower_template) != best_period: + state.lower_template = resample_template( + np.array(state.lower_template, dtype=float), best_period + ).astype(float).tolist() + + if len(state.upper_template) != best_period: + state.upper_template = resample_template( + np.array(state.upper_template, dtype=float), best_period + ).astype(float).tolist() + + state.period = best_period + state.phase_origin_ts = int(best_origin) + + +def maybe_update_state( + key: str, + ts_grid: np.ndarray, + ys_model: np.ndarray, + ys_actual: np.ndarray, + target: Dict, + states: Dict[str, BaselineState], +) -> Tuple[Optional[BaselineState], bool, float, float, float, int, float]: + """ + 核心状态更新函数：检测异常并按状态机规则更新基线。 + + 状态机转换： + - 无状态 → 初始化 → HEALTHY（返回，本轮不做异常检测） + - HEALTHY + 异常 → ANOMALY（冻结模板） + - ANOMALY + 正常 → RECOVERING（开始计时） + - RECOVERING + 正常 + 足够时间 → HEALTHY（恢复学习） + - HEALTHY/RECOVERING + 正常 + 足够时间 → 更新模板（EMA） + + Args: + key: 序列唯一标识符（用于 states 字典的键） + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_model: 平滑后的信号 + ys_actual: 原始信号 + target: target dict，包含策略和阈值配置 + states: 所有指标的状态字典（由 PredictorService 持有，原地修改） + + Returns: + (state, is_anomaly, outside_ratio, mean_abs_err, mean_rel_err, + max_outside_seconds, max_exceed_ratio) + state 为 None 表示数据不足，本轮跳过。 + """ + now_sec = int(time.time()) + state = states.get(key) + + # 首次见到该指标：初始化健康基线 + if state is None: + state = create_initial_state( + ts_grid=ts_grid, + ys_model=ys_model, + ys_actual=ys_actual, + target=target, + now_sec=now_sec, + ) + + if state is None: + return None, False, 0.0, 0.0, 0.0, 0, 0.0 + + states[key] = state + logger.info( + "初始化健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss", + key, + state.strategy, + state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, + ) + return state, False, 0.0, 0.0, 0.0, 0, 0.0 + + # 计算距上次处理的时间（秒），用于累加 clean_seconds + elapsed = max(1, now_sec - int(state.last_seen_ts)) + elapsed = min(elapsed, config.POLL_INTERVAL * 2) # 防止长时间停机后 clean_seconds 暴增 + state.last_seen_ts = now_sec + + ( + is_anomaly, + outside_ratio, + mean_abs_err, + mean_rel_err, + best_period, + best_origin, + max_outside_seconds, + max_exceed_ratio, + ) = detect_anomaly( + state=state, + ts_grid=ts_grid, + ys_model=ys_model, + ys_actual=ys_actual, + target=target, + ) + + # 异常：冻结模板，不学习故障数据 + if is_anomaly: + state.status = BASELINE_STATUS_ANOMALY + state.clean_seconds = 0 + states[key] = state + logger.warning( + "检测到异常，冻结模板 key=%s outside_ratio=%.2f max_outside=%ss " + "max_exceed_ratio=%.2f mean_abs_err=%.4f mean_rel_err=%.4f", + key, outside_ratio, max_outside_seconds, + max_exceed_ratio, mean_abs_err, mean_rel_err, + ) + return state, True, outside_ratio, mean_abs_err, mean_rel_err, max_outside_seconds, max_exceed_ratio + + # 正常：应用 phase-lock 结果 + old_period = int(state.period) + old_origin = int(state.phase_origin_ts) + apply_phase_lock_to_state(state, best_period, best_origin) + + if old_period != state.period or old_origin != state.phase_origin_ts: + logger.info( + "phase-lock key=%s period %s -> %s origin %s -> %s", + key, old_period, state.period, + datetime.fromtimestamp(old_origin).strftime("%H:%M:%S"), + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + ) + + # 异常刚消失：进入恢复期，等待稳定后再恢复学习 + if state.status == BASELINE_STATUS_ANOMALY: + state.status = BASELINE_STATUS_RECOVERING + state.clean_seconds = elapsed + states[key] = state + logger.info("异常开始恢复 key=%s clean_seconds=%ss", key, state.clean_seconds) + return state, False, outside_ratio, mean_abs_err, mean_rel_err, max_outside_seconds, max_exceed_ratio + + # 累加健康时间 + if state.status == BASELINE_STATUS_RECOVERING: + state.clean_seconds += elapsed + else: + state.status = BASELINE_STATUS_HEALTHY + state.clean_seconds += elapsed + + # 健康时间不足：不更新模板 + min_clean_for_update = max( + config.RECOVERY_MIN_SECONDS, + int(state.period) * config.MIN_FULL_CYCLES_FOR_TEMPLATE, + ) + if state.clean_seconds < min_clean_for_update: + states[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err, max_outside_seconds, max_exceed_ratio + + # 健康时间足够：用最近数据更新模板（EMA） + tail_seconds = min( + int(state.clean_seconds), + int(state.period) * config.MAX_CYCLES_FOR_TEMPLATE, + ) + + strategy = str(target.get("strategy", "phase_point")) + band_low_q = float(target.get("band_low_q", 5.0)) + band_high_q = float(target.get("band_high_q", 95.0)) + + baseline = build_current_baseline( + ts_grid=ts_grid, + ys_mid_grid=ys_model, + ys_band_grid=ys_actual, + strategy=strategy, + band_low_q=band_low_q, + band_high_q=band_high_q, + tail_seconds=tail_seconds, + ) + + if baseline is None: + states[key] = state + return state, False, outside_ratio, mean_abs_err, mean_rel_err, max_outside_seconds, max_exceed_ratio + + new_period, new_origin, new_template, new_lower_template, new_upper_template = baseline + + # 恢复期用更激进的 alpha，加速追赶真实信号 + alpha = config.RECOVERY_EMA_ALPHA if state.status == BASELINE_STATUS_RECOVERING else config.HEALTHY_EMA_ALPHA + + state.template = merge_template( + np.array(state.template, dtype=float), new_template, alpha + ).astype(float).tolist() + state.lower_template = merge_template( + np.array(state.lower_template, dtype=float), new_lower_template, alpha + ).astype(float).tolist() + state.upper_template = merge_template( + np.array(state.upper_template, dtype=float), new_upper_template, alpha + ).astype(float).tolist() + + state.period = int(new_period) + state.phase_origin_ts = int(new_origin) + state.status = BASELINE_STATUS_HEALTHY + state.last_update_ts = now_sec + + # 更新量程统计（用于 Grafana 展示） + if tail_seconds > 0 and len(ys_actual) >= tail_seconds: + state.y_min = float(np.min(ys_actual[-tail_seconds:])) + state.y_max = float(np.max(ys_actual[-tail_seconds:])) + else: + state.y_min = float(np.min(ys_actual)) + state.y_max = float(np.max(ys_actual)) + + states[key] = state + logger.info( + "更新健康模板 key=%s strategy=%s period=%ss origin=%s clean=%ss alpha=%.2f", + key, state.strategy, state.period, + datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S"), + state.clean_seconds, alpha, + ) + + return state, False, outside_ratio, mean_abs_err, mean_rel_err, max_outside_seconds, max_exceed_ratio diff --git a/ai/predictor/storage.py b/ai/predictor/storage.py new file mode 100644 index 0000000..f614528 --- /dev/null +++ b/ai/predictor/storage.py @@ -0,0 +1,438 @@ +# -*- coding: utf-8 -*- +""" +predictor.storage +~~~~~~~~~~~~~~~~~ +VictoriaMetrics 读写层，封装所有网络 IO。 + +职责： +- 从 VM 拉取历史时序数据（query_range） +- 将预测结果和异常指标写入 VM（import/prometheus） +- 标签字符串的序列化与解析 +- 状态文件的持久化读写 + +本模块不包含任何预测或异常检测逻辑，只负责数据的搬运和格式转换。 + +依赖：requests, numpy +""" + +import json +import logging +import math +import os +import re +from dataclasses import asdict +from datetime import datetime, timedelta +from typing import Dict, List, Optional, Tuple + +import numpy as np +import requests + +from .models import BaselineState + +logger = logging.getLogger(__name__) + +# 用于解析 PromQL 标签字符串的正则，匹配 key="value" 格式 +_LABEL_PATTERN = re.compile( + r'\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*=\s*"((?:\\.|[^"])*)"\s*' +) + + +# --------------------------------------------------------------------------- +# 历史数据读取 +# --------------------------------------------------------------------------- + +def fetch_history( + vm_url: str, + query: str, + minutes: int, + step: str = "1s", +) -> Tuple[List[float], List[float]]: + """ + 从 VictoriaMetrics 拉取指定查询的历史时序数据。 + + Args: + vm_url: VM HTTP 地址，如 "http://localhost:8428" + query: PromQL 查询表达式，如 'feed_rate{device_id="fanuc-cnc"}' + minutes: 向前拉取多少分钟的历史数据 + step: 查询步长，默认 "1s"（每秒一个点） + + Returns: + (timestamps, values) 两个列表，长度相同。 + 如果查询失败或无数据，返回两个空列表。 + """ + now = datetime.now() + start = now - timedelta(minutes=minutes) + + try: + resp = requests.get( + f"{vm_url}/api/v1/query_range", + params={ + "query": query, + "start": start.timestamp(), + "end": now.timestamp(), + "step": step, + }, + timeout=10, + ) + resp.raise_for_status() + except requests.RequestException as e: + logger.error("拉取历史数据失败 query=%s: %s", query, e) + return [], [] + + try: + result = resp.json().get("data", {}).get("result", []) + except Exception as e: + logger.error("解析 VM 响应失败 query=%s: %s", query, e) + return [], [] + + if not result: + return [], [] + + ts_list: List[float] = [] + ys_list: List[float] = [] + + for item in result[0].get("values", []): + if len(item) < 2: + continue + try: + t = float(item[0]) + y = float(item[1]) + except (TypeError, ValueError): + continue + # 过滤 NaN / Inf，防止后续 numpy 计算出错 + if math.isfinite(t) and math.isfinite(y): + ts_list.append(t) + ys_list.append(y) + + return ts_list, ys_list + + +def normalize_history( + ts: List[float], + ys: List[float], +) -> Tuple[np.ndarray, np.ndarray]: + """ + 将原始时序数据规整化为均匀 1 秒间隔的网格。 + + 处理步骤： + 1. 去重（同一秒内取最后一个值） + 2. 按时间戳排序 + 3. 线性插值填充缺失秒 + + Args: + ts: 原始时间戳列表（Unix 秒，可以是浮点数） + ys: 对应的值列表 + + Returns: + (ts_grid, ys_grid) 均匀 1 秒间隔的 numpy 数组。 + 如果输入无效，返回两个空数组。 + """ + if not ts or not ys or len(ts) != len(ys): + return np.array([]), np.array([]) + + # 去重：同一秒取最后写入的值 + data: Dict[int, float] = {} + for t, y in zip(ts, ys): + try: + sec = int(round(float(t))) + val = float(y) + except (TypeError, ValueError): + continue + if math.isfinite(sec) and math.isfinite(val): + data[sec] = val + + if not data: + return np.array([]), np.array([]) + + sorted_items = sorted(data.items()) + ts_clean = np.array([x[0] for x in sorted_items], dtype=float) + ys_clean = np.array([x[1] for x in sorted_items], dtype=float) + + if len(ts_clean) < 2: + return ts_clean, ys_clean + + start_sec = int(ts_clean[0]) + end_sec = int(ts_clean[-1]) + + if end_sec <= start_sec: + return ts_clean, ys_clean + + # 构建均匀网格并插值 + ts_grid = np.arange(start_sec, end_sec + 1, 1, dtype=float) + ys_grid = np.interp(ts_grid, ts_clean, ys_clean) + + return ts_grid, ys_grid + + +# --------------------------------------------------------------------------- +# 标签工具 +# --------------------------------------------------------------------------- + +def prom_escape_label_value(value: str) -> str: + """对 Prometheus 标签值进行转义，处理反斜杠、换行符和双引号。""" + return ( + str(value) + .replace("\\", "\\\\") + .replace("\n", "\\n") + .replace('"', '\\"') + ) + + +def labels_to_str(labels: Dict[str, str]) -> str: + """ + 将标签字典序列化为 Prometheus 格式的标签字符串。 + + Example: + {"device_id": "fanuc-cnc", "source": "protoforge"} + → '{device_id="fanuc-cnc",source="protoforge"}' + """ + if not labels: + return "" + parts = [ + f'{k}="{prom_escape_label_value(labels[k])}"' + for k in sorted(labels) + ] + return "{" + ",".join(parts) + "}" + + +def parse_labels_from_query(query: str) -> Dict[str, str]: + """ + 从 PromQL 查询字符串中提取标签字典。 + + Example: + 'feed_rate{device_id="fanuc-cnc"}' → {"device_id": "fanuc-cnc"} + """ + labels: Dict[str, str] = {} + + if "{" not in query or "}" not in query: + return labels + + try: + label_part = query[query.index("{") + 1 : query.rindex("}")] + except ValueError: + return labels + + for match in _LABEL_PATTERN.finditer(label_part): + key = match.group(1) + value = ( + match.group(2) + .replace('\\"', '"') + .replace("\\n", "\n") + .replace("\\\\", "\\") + ) + labels[key] = value + + return labels + + +def merge_labels(*dicts: Dict[str, str]) -> Dict[str, str]: + """合并多个标签字典，后面的字典覆盖前面的同名键。""" + result: Dict[str, str] = {} + for d in dicts: + if d: + result.update(d) + return result + + +def series_key(metric_name: str, labels: Dict[str, str]) -> str: + """生成唯一的序列标识符，用于 BaselineState 字典的键。""" + return metric_name + labels_to_str(labels) + + +# --------------------------------------------------------------------------- +# 数据写入 +# --------------------------------------------------------------------------- + +def write_series( + vm_url: str, + metric_name: str, + labels: Dict[str, str], + ts_list: List[int], + values: List[float], +) -> bool: + """ + 将一条时序数据写入 VictoriaMetrics（Prometheus remote write 格式）。 + + Args: + vm_url: VM HTTP 地址 + metric_name: 指标名 + labels: 标签字典 + ts_list: 时间戳列表（Unix 秒） + values: 对应的值列表 + + Returns: + 写入成功返回 True，否则返回 False。 + """ + if not ts_list or not values or len(ts_list) != len(values): + return False + + label_str = labels_to_str(labels) + lines: List[str] = [] + + for t, y in zip(ts_list, values): + try: + ts_sec = int(round(float(t))) + val = float(y) + except (TypeError, ValueError): + continue + if math.isfinite(ts_sec) and math.isfinite(val): + # VM 使用毫秒时间戳 + lines.append(f"{metric_name}{label_str} {val:.6f} {ts_sec * 1000}") + + if not lines: + return False + + payload = "\n".join(lines) + "\n" + + try: + resp = requests.post( + f"{vm_url}/api/v1/import/prometheus", + data=payload.encode("utf-8"), + headers={"Content-Type": "text/plain; version=0.0.4; charset=utf-8"}, + timeout=10, + ) + resp.raise_for_status() + return True + except requests.RequestException as e: + logger.error("写入数据失败 metric=%s: %s", metric_name, e) + return False + + +def write_prediction_bundle( + vm_url: str, + pred_metric: str, + anomaly_metric: str, + labels: Dict[str, str], + ts_future: List[int], + pred_values: np.ndarray, + lower_values: np.ndarray, + upper_values: np.ndarray, + is_anomaly: bool, + outside_ratio: float, + mean_abs_err: float, + mean_rel_err: float, + max_outside_seconds: int, + max_exceed_ratio: float, + event_ts: int, +) -> bool: + """ + 一次性写入一个指标的完整预测结果包，包含： + - 预测中值曲线（pred_metric） + - 预测下界曲线（pred_metric_lower） + - 预测上界曲线（pred_metric_upper） + - 异常标志（anomaly_metric，0 或 1） + - 各项异常诊断指标（outside_ratio、mean_abs_error 等） + + Args: + vm_url: VM HTTP 地址 + pred_metric: 预测指标名，如 "feed_rate_predicted" + anomaly_metric: 异常指标名，如 "feed_rate_anomaly" + labels: 写入时附加的标签 + ts_future: 预测时间戳列表（未来时刻，Unix 秒） + pred_values: 预测中值数组 + lower_values: 预测下界数组 + upper_values: 预测上界数组 + is_anomaly: 当前是否判定为异常 + outside_ratio: 检测窗口内越界点比例 + mean_abs_err: 平均绝对误差 + mean_rel_err: 平均相对误差 + max_outside_seconds: 最长连续越界秒数 + max_exceed_ratio: 最大越界倍数（相对于边界宽度） + event_ts: 异常诊断指标的时间戳（通常为最后一个真实数据点的时间戳） + + Returns: + 所有写入均成功返回 True，任意一个失败返回 False。 + """ + # 异常诊断指标附加 type 标签，便于在 Grafana 中过滤 + anomaly_labels = {**labels, "type": "prediction_deviation"} + + results = [ + write_series(vm_url, pred_metric, labels, + ts_future, pred_values.tolist()), + write_series(vm_url, f"{pred_metric}_lower", labels, + ts_future, lower_values.tolist()), + write_series(vm_url, f"{pred_metric}_upper", labels, + ts_future, upper_values.tolist()), + write_series(vm_url, anomaly_metric, anomaly_labels, + [event_ts], [1.0 if is_anomaly else 0.0]), + write_series(vm_url, f"{anomaly_metric}_outside_ratio", anomaly_labels, + [event_ts], [outside_ratio]), + write_series(vm_url, f"{anomaly_metric}_mean_abs_error", anomaly_labels, + [event_ts], [mean_abs_err]), + write_series(vm_url, f"{anomaly_metric}_mean_rel_error", anomaly_labels, + [event_ts], [mean_rel_err]), + write_series(vm_url, f"{anomaly_metric}_max_consecutive_outside", anomaly_labels, + [event_ts], [float(max_outside_seconds)]), + write_series(vm_url, f"{anomaly_metric}_max_exceed_ratio", anomaly_labels, + [event_ts], [float(max_exceed_ratio)]), + ] + + return all(results) + + +# --------------------------------------------------------------------------- +# 状态持久化 +# --------------------------------------------------------------------------- + +def load_state(path: str) -> Dict[str, BaselineState]: + """ + 从 JSON 文件加载所有指标的基线状态。 + + 文件不存在时返回空字典（正常首次启动情况）。 + 字段不完整的条目会被跳过，不会导致整体加载失败。 + + Args: + path: 状态文件路径 + + Returns: + key → BaselineState 的字典 + """ + if not os.path.exists(path): + return {} + + try: + with open(path, "r", encoding="utf-8") as f: + raw = json.load(f) + except Exception as e: + logger.warning("加载状态文件失败，将重新学习: %s", e) + return {} + + required_fields = { + "period", "phase_origin_ts", "template", "lower_template", + "upper_template", "strategy", "status", "clean_seconds", + "last_update_ts", "last_seen_ts", "y_min", "y_max", + } + + states: Dict[str, BaselineState] = {} + for key, value in raw.get("baseline_states", {}).items(): + if required_fields.issubset(value.keys()): + states[key] = BaselineState(**value) + + logger.info("已加载状态文件 %s，共 %d 条记录", path, len(states)) + return states + + +def save_state(path: str, states: Dict[str, BaselineState]) -> None: + """ + 将所有指标的基线状态原子写入 JSON 文件。 + + 使用临时文件 + os.replace 保证写入原子性， + 避免进程崩溃时产生损坏的状态文件。 + + Args: + path: 状态文件路径 + states: key → BaselineState 的字典 + """ + try: + raw = { + "baseline_states": { + key: asdict(state) + for key, state in states.items() + } + } + tmp_path = path + ".tmp" + with open(tmp_path, "w", encoding="utf-8") as f: + json.dump(raw, f, ensure_ascii=False, indent=2) + os.replace(tmp_path, path) + except Exception as e: + logger.warning("保存状态文件失败: %s", e) diff --git a/ai/predictor/template.py b/ai/predictor/template.py new file mode 100644 index 0000000..86d8170 --- /dev/null +++ b/ai/predictor/template.py @@ -0,0 +1,384 @@ +# -*- coding: utf-8 -*- +""" +predictor.template +~~~~~~~~~~~~~~~~~~ +模板的构建、预测、重采样与融合，不包含任何 IO 操作。 + +职责： +- 从历史谷底片段构建周期模板（中值/分位数） +- 基于模板和相位原点预测未来值 +- 模板重采样（周期变化时对齐长度） +- EMA 融合新旧模板（渐进式学习） +- 相位原点规整化 + +依赖：numpy, predictor.signal, predictor.config, predictor.models +""" + +import math +from typing import List, Optional, Tuple + +import numpy as np + +from . import config +from .models import BaselineState +from .signal import moving_average + + +# --------------------------------------------------------------------------- +# 模板构建 +# --------------------------------------------------------------------------- + +def build_templates_from_valleys( + ts_grid: np.ndarray, + ys_mid_grid: np.ndarray, + ys_band_grid: np.ndarray, + period: int, + valleys: List[int], + strategy: str, + band_low_q: float, + band_high_q: float, +) -> Optional[Tuple[np.ndarray, np.ndarray, np.ndarray]]: + """ + 从历史谷底片段构建三条模板曲线（中值、下界、上界）。 + + 每个相邻谷底对定义一个周期片段，将其重采样到统一的 period 长度， + 再按策略聚合： + - phase_point：加权平均（越近的周期权重越高） + - phase_band：中位数 + 分位数（对异常周期鲁棒） + + Args: + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_mid_grid: 用于构建中值模板的平滑信号（phase_point 用） + ys_band_grid: 用于构建分位数模板的原始信号（phase_band 用） + period: 目标模板长度（秒） + valleys: 谷底索引列表 + strategy: "phase_point" 或 "phase_band" + band_low_q: phase_band 下界分位数（如 5.0） + band_high_q: phase_band 上界分位数（如 95.0） + + Returns: + (mid_template, lower_template, upper_template) 三个长度为 period 的数组。 + 数据不足时返回 None。 + """ + if period <= 1 or len(valleys) < config.MIN_FULL_CYCLES_FOR_TEMPLATE + 1: + return None + + # 筛选长度合理的周期片段（0.55~1.60 倍期望周期） + pairs = [ + (a, b, float(ts_grid[b] - ts_grid[a])) + for a, b in zip(valleys[:-1], valleys[1:]) + if period * 0.55 <= float(ts_grid[b] - ts_grid[a]) <= period * 1.60 + ] + + if len(pairs) < config.MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + # 只取最近 MAX_CYCLES_FOR_TEMPLATE 个周期，防止过旧数据污染 + pairs = pairs[-config.MAX_CYCLES_FOR_TEMPLATE:] + + phase_grid = np.arange(period, dtype=float) + mid_segments: List[np.ndarray] = [] + band_segments: List[np.ndarray] = [] + weights: List[float] = [] + + for idx, (a, b, cycle_len) in enumerate(pairs): + seg_ts = ts_grid[a : b + 1] + seg_mid_y = ys_mid_grid[a : b + 1] + seg_band_y = ys_band_grid[a : b + 1] + + if len(seg_mid_y) < 3 or len(seg_band_y) < 3: + continue + + # 将片段的时间轴归一化到 [0, period)，再插值到统一相位网格 + x_old = (seg_ts - seg_ts[0]) / cycle_len * period + mid_seg = np.interp(phase_grid, x_old, seg_mid_y) + band_seg = np.interp(phase_grid, x_old, seg_band_y) + + mid_segments.append(mid_seg.astype(float)) + band_segments.append(band_seg.astype(float)) + # 越近的周期权重越高（线性递增，范围 0.5~1.0） + weights.append(0.5 + 0.5 * ((idx + 1) / len(pairs))) + + if len(mid_segments) < config.MIN_FULL_CYCLES_FOR_TEMPLATE: + return None + + mid_arr = np.vstack(mid_segments) + band_arr = np.vstack(band_segments) + w_arr = np.array(weights, dtype=float) + + if strategy == "phase_band": + # 分位数聚合：对异常周期鲁棒，保留正常波动范围 + mid_template = np.percentile(mid_arr, 50, axis=0) + lower_template = np.percentile(band_arr, band_low_q, axis=0) + upper_template = np.percentile(band_arr, band_high_q, axis=0) + else: + # 加权平均：越近的周期贡献越大 + mid_template = np.average(mid_arr, axis=0, weights=w_arr) + lower_template = mid_template.copy() + upper_template = mid_template.copy() + + return ( + mid_template.astype(float), + lower_template.astype(float), + upper_template.astype(float), + ) + + +def build_current_baseline( + ts_grid: np.ndarray, + ys_mid_grid: np.ndarray, + ys_band_grid: np.ndarray, + strategy: str, + band_low_q: float, + band_high_q: float, + tail_seconds: Optional[int] = None, +) -> Optional[Tuple[int, int, np.ndarray, np.ndarray, np.ndarray]]: + """ + 从历史数据构建当前基线（周期 + 相位原点 + 三条模板曲线）。 + + 可选 tail_seconds 参数限制只使用最近一段数据， + 用于健康状态下的增量模板更新（避免使用过旧的异常数据）。 + + Args: + ts_grid: 均匀 1 秒网格的时间戳数组 + ys_mid_grid: 平滑后的信号（用于周期检测和中值模板） + ys_band_grid: 原始信号（用于分位数模板） + strategy: "phase_point" 或 "phase_band" + band_low_q: phase_band 下界分位数 + band_high_q: phase_band 上界分位数 + tail_seconds: 若指定，只使用最近 tail_seconds 秒的数据 + + Returns: + (period, phase_origin_ts, template, lower_template, upper_template) + 数据不足或无法检测到谷底时返回 None。 + """ + from .signal import detect_period_and_valleys + + if len(ys_mid_grid) < config.MIN_POINTS or len(ys_band_grid) < config.MIN_POINTS: + return None + + if tail_seconds is not None and tail_seconds > 0: + cutoff = ts_grid[-1] - int(tail_seconds) + mask = ts_grid >= cutoff + ts_use = ts_grid[mask] + ys_mid_use = ys_mid_grid[mask] + ys_band_use = ys_band_grid[mask] + else: + ts_use = ts_grid + ys_mid_use = ys_mid_grid + ys_band_use = ys_band_grid + + if len(ys_mid_use) < config.MIN_POINTS or len(ys_band_use) < config.MIN_POINTS: + return None + + period, valleys = detect_period_and_valleys(ts_use, ys_mid_use) + + templates = build_templates_from_valleys( + ts_grid=ts_use, + ys_mid_grid=ys_mid_use, + ys_band_grid=ys_band_use, + period=period, + valleys=valleys, + strategy=strategy, + band_low_q=band_low_q, + band_high_q=band_high_q, + ) + + if templates is None or len(valleys) == 0: + return None + + template, lower_template, upper_template = templates + # 以最后一个谷底作为相位原点 + phase_origin_ts = int(round(float(ts_use[valleys[-1]]))) + + return int(period), phase_origin_ts, template, lower_template, upper_template + + +# --------------------------------------------------------------------------- +# 模板预测 +# --------------------------------------------------------------------------- + +def circular_template_value(template: np.ndarray, phase: float) -> float: + """ + 从模板中读取指定相位处的值（线性插值，循环边界）。 + + Args: + template: 长度为 period 的模板数组 + phase: 相位（0 到 period 之间的浮点数） + + Returns: + 插值后的模板值。 + """ + period = len(template) + if period == 0: + return 0.0 + + phase = float(phase) % period + i0 = int(math.floor(phase)) % period + i1 = (i0 + 1) % period + frac = phase - math.floor(phase) + + return float((1.0 - frac) * template[i0] + frac * template[i1]) + + +def resample_template(old_template: np.ndarray, new_period: int) -> np.ndarray: + """ + 将模板重采样到新的周期长度。 + + 当 phase-lock 检测到周期漂移时，需要将旧模板拉伸/压缩到新周期。 + 使用循环扩展（拼接三份）保证边界处插值正确。 + + Args: + old_template: 原始模板数组 + new_period: 目标周期长度（秒） + + Returns: + 重采样后的模板数组，长度为 new_period。 + """ + old_period = len(old_template) + if old_period == new_period: + return old_template.astype(float) + + if old_period <= 1 or new_period <= 1: + return np.full(new_period, float(np.mean(old_template)), dtype=float) + + # 归一化到 [0, 1) 相位空间，循环扩展保证边界插值正确 + old_x = np.linspace(0.0, 1.0, old_period, endpoint=False) + new_x = np.linspace(0.0, 1.0, new_period, endpoint=False) + + old_x_ext = np.concatenate([old_x - 1.0, old_x, old_x + 1.0]) + old_y_ext = np.concatenate([old_template, old_template, old_template]) + + return np.interp(new_x, old_x_ext, old_y_ext).astype(float) + + +def predict_template_values( + template: np.ndarray, + period: int, + phase_origin_ts: int, + ts_list: List[int], +) -> np.ndarray: + """ + 根据模板和相位原点，预测一组时间戳处的值。 + + 相位 = (ts - phase_origin_ts) mod period， + 再从模板中线性插值读取对应值。 + + Args: + template: 长度为 period 的模板数组 + period: 周期（秒） + phase_origin_ts: 相位原点时间戳（Unix 秒） + ts_list: 待预测的时间戳列表（Unix 秒） + + Returns: + 预测值数组，长度与 ts_list 相同。 + """ + if period <= 1: + return np.zeros(len(ts_list), dtype=float) + + if len(template) != period: + template = resample_template(template, period) + + values = [ + circular_template_value(template, (int(ts) - int(phase_origin_ts)) % period) + for ts in ts_list + ] + return np.array(values, dtype=float) + + +def predict_state_bundle( + state: BaselineState, + ts_list: List[int], +) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: + """ + 用 BaselineState 中的三条模板预测一组时间戳处的值。 + + Args: + state: 当前基线状态（包含 period、phase_origin_ts、三条模板） + ts_list: 待预测的时间戳列表（Unix 秒） + + Returns: + (mid, lower, upper) 三个预测数组，长度与 ts_list 相同。 + """ + period = int(state.period) + origin = int(state.phase_origin_ts) + + mid = predict_template_values( + template=np.array(state.template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + lower = predict_template_values( + template=np.array(state.lower_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + upper = predict_template_values( + template=np.array(state.upper_template, dtype=float), + period=period, + phase_origin_ts=origin, + ts_list=ts_list, + ) + return mid, lower, upper + + +def normalize_origin_near(origin: int, period: int, near_ts: int) -> int: + """ + 将相位原点规整化到 near_ts 附近（使 origin <= near_ts < origin + period）。 + + phase-lock 搜索时需要将原点移到最近的时间窗口内， + 避免因原点过旧导致相位计算溢出。 + + Args: + origin: 当前相位原点（Unix 秒） + period: 周期（秒） + near_ts: 目标时间戳（通常为最新数据点的时间戳） + + Returns: + 规整化后的相位原点（Unix 秒）。 + """ + if period <= 1: + return origin + + origin = int(origin) + period = int(period) + near_ts = int(near_ts) + + while origin + period <= near_ts: + origin += period + + while origin > near_ts: + origin -= period + + return origin + + +def merge_template( + old_template: np.ndarray, + new_template: np.ndarray, + alpha: float, +) -> np.ndarray: + """ + 用 EMA 融合旧模板和新模板。 + + merged = (1 - alpha) * old + alpha * new + + 若两者长度不同，先将旧模板重采样到新模板长度。 + alpha 越大，新模板权重越高（学习越激进）。 + + Args: + old_template: 旧模板数组 + new_template: 新模板数组 + alpha: EMA 步长，clip 到 [0, 1] + + Returns: + 融合后的模板数组，长度与 new_template 相同。 + """ + alpha = float(np.clip(alpha, 0.0, 1.0)) + + if len(old_template) != len(new_template): + old_template = resample_template(old_template, len(new_template)) + + return ((1.0 - alpha) * old_template + alpha * new_template).astype(float) diff --git a/ai/pridict_v5.py b/ai/pridict_v5.py index 6894a66..dde0b11 100644 --- a/ai/pridict_v5.py +++ b/ai/pridict_v5.py @@ -1,27 +1,31 @@ # -*- coding: utf-8 -*- """ -ProtoForge Predictor v12 +ProtoForge Predictor v13 核心能力： -1. feed_rate / spindle_speed / spindle_current 使用 phase-lock 点预测。 -2. vibration_x / vibration_y / vibration_z 使用 phase-band 预测带。 -3. vibration 类指标： - - predicted 使用平滑后的中位数模板，用于趋势参考。 - - upper/lower 使用原始波动分位数模板 + padding，用于正常波动容忍带。 - - 偶发越界不直接报警，只有持续越界 / 高比例越界 / 严重越界才报警。 -4. 预测起点锚定最后一个真实点 last_real_ts，避免时间错位。 -5. 异常期间冻结健康模板，不学习故障数据。 -6. 故障恢复后等待稳定，再恢复模板学习。 -7. 写入： - - xxx_predicted - - xxx_predicted_upper - - xxx_predicted_lower - - xxx_anomaly - - xxx_anomaly_outside_ratio - - xxx_anomaly_mean_abs_error - - xxx_anomaly_mean_rel_error - - xxx_anomaly_max_consecutive_outside - - xxx_anomaly_max_exceed_ratio +1. 支持三个独立 CNC 工位：粗铣(fanuc-cnc)、半精铣(fanuc-cnc-semi-finish)、精铣(fanuc-cnc-finish) +2. 覆盖指标：feed_rate / spindle_speed / spindle_current / spindle_load +3. feed_rate / spindle_speed / spindle_current 使用 phase-lock 点预测。 +4. spindle_load 使用 phase_band 预测带（多频漂移容忍）。 +5. vibration_x / vibration_y / vibration_z 使用 phase-band 预测带。 +6. 各工位独立阈值配置，匹配实际量程差异： + - 粗铣：spindle_speed~2000RPM, feed_rate~800mm/min, spindle_current~21A, spindle_load~56% + - 半精铣：spindle_speed~4000RPM, feed_rate~500mm/min, spindle_current~14.5A, spindle_load~38% + - 精铣：spindle_speed~6000RPM, feed_rate~300mm/min, spindle_current~8.5A, spindle_load~22% +7. 粗铣周期含随机抖动(±10s)，phase-lock 搜索范围扩大至 ±18%。 +8. 预测起点锚定最后一个真实点 last_real_ts，避免时间错位。 +9. 异常期间冻结健康模板，不学习故障数据。 +10. 故障恢复后等待稳定，再恢复模板学习。 +11. 写入： + - xxx_predicted + - xxx_predicted_upper + - xxx_predicted_lower + - xxx_anomaly + - xxx_anomaly_outside_ratio + - xxx_anomaly_mean_abs_error + - xxx_anomaly_mean_rel_error + - xxx_anomaly_max_consecutive_outside + - xxx_anomaly_max_exceed_ratio """ import json @@ -55,7 +59,7 @@ # ============================================================================= VM_URL = "http://localhost:8428" -STATE_FILE = "/tmp/protoforge_predictor_state_v12.json" +STATE_FILE = "/tmp/protoforge_predictor_state_v14.json" HISTORY_MINUTES = 30 HORIZON_SECONDS = 120 @@ -86,6 +90,7 @@ MAX_DATA_LAG_SECONDS = 180 +# 默认 phase-lock 搜索参数（精铣/半精铣：固定周期，搜索范围窄） PHASE_LOCK_MIN_WINDOW_SECONDS = 45 PHASE_LOCK_MAX_WINDOW_SECONDS = 180 PHASE_LOCK_PERIOD_SEARCH_RATIO = 0.12 @@ -95,105 +100,260 @@ # ============================================================================= -# 指标配置 +# 监控指标白名单（可通过环境变量 PROTOFORGE_MONITORED_METRICS 覆盖） # ============================================================================= -PREDICT_TARGETS = [ - { - "query": 'feed_rate{device_id="fanuc-cnc"}', - "pred_metric": "feed_rate_predicted", - "anomaly_metric": "feed_rate_anomaly", - "strategy": "phase_point", - "abs_threshold": 400.0, - "rel_threshold": 0.25, - "smooth_window": 1, - "outside_ratio_threshold": 0.60, - "min_consecutive_outside": 5, - "severe_exceed_ratio": 1.8, - }, - { - "query": 'spindle_speed{device_id="fanuc-cnc"}', - "pred_metric": "spindle_speed_predicted", - "anomaly_metric": "spindle_speed_anomaly", - "strategy": "phase_point", - "abs_threshold": 500.0, - "rel_threshold": 0.25, - "smooth_window": 1, - "outside_ratio_threshold": 0.60, - "min_consecutive_outside": 5, - "severe_exceed_ratio": 1.8, - }, - { - "query": 'spindle_current{device_id="fanuc-cnc"}', - "pred_metric": "spindle_current_predicted", - "anomaly_metric": "spindle_current_anomaly", - "strategy": "phase_point", - "abs_threshold": 5.0, - "rel_threshold": 0.25, - "smooth_window": 1, +_DEFAULT_MONITORED_METRICS = [ + "feed_rate", + "spindle_speed", + "spindle_current", + "spindle_load", + "vibration_x", + "vibration_y", + "vibration_z", +] + +MONITORED_METRICS: List[str] = [ + m.strip() + for m in os.environ.get( + "PROTOFORGE_MONITORED_METRICS", + ",".join(_DEFAULT_MONITORED_METRICS), + ).split(",") + if m.strip() +] + +# 人工上下限覆盖文件（可选，不存在则忽略） +# 格式：{"device-id": {"metric_name": {"hard_max": 35.0, "hard_min": 0.0}}} +OVERRIDE_FILE = os.environ.get( + "PROTOFORGE_PREDICTOR_OVERRIDE", + "/etc/protoforge/predictor_override.json", +) + +# 目标列表刷新间隔（秒） +TARGETS_REFRESH_INTERVAL = int(os.environ.get("PROTOFORGE_TARGETS_REFRESH", "60")) + +# 运行时目标缓存 +_TARGETS_CACHE: List[Dict] = [] +_TARGETS_LAST_REFRESH: float = 0.0 + + +# ============================================================================= +# Layer 1: 设备与指标发现 +# ============================================================================= + +def discover_device_ids() -> List[str]: + """查询 VM 中所有 device_id 标签值。""" + try: + resp = requests.get( + f"{VM_URL}/api/v1/label/device_id/values", + timeout=10, + ) + resp.raise_for_status() + return [v for v in resp.json().get("data", []) if v] + except requests.RequestException as e: + logger.error("发现 device_id 失败: %s", e) + return [] + + +def discover_metrics_for_device(device_id: str) -> List[str]: + """查询该设备在 VM 中实际存在且有近期数据的指标名。""" + found = [] + for metric in MONITORED_METRICS: + try: + resp = requests.get( + f"{VM_URL}/api/v1/query", + params={"query": f'{metric}{{device_id="{device_id}"}}'}, + timeout=5, + ) + resp.raise_for_status() + if resp.json().get("data", {}).get("result"): + found.append(metric) + except requests.RequestException: + pass + return found + + +# ============================================================================= +# Layer 2: 自适应配置推断 +# ============================================================================= + +@dataclass +class MetricProfile: + """从历史数据统计出的指标特征，驱动策略和阈值的自动推断。""" + device_id: str + metric: str + p5: float # 活跃段 5th percentile + p95: float # 活跃段 95th percentile + iqr: float # p95 - p5 + cv: float # 变异系数 std/mean（衡量稳定性） + strategy: str # "phase_point" 或 "phase_band" + abs_threshold: float + rel_threshold: float + band_low_q: float + band_high_q: float + band_pad_abs: float + phase_lock_period_search_ratio: float + + +def infer_metric_profile(device_id: str, metric: str) -> Optional["MetricProfile"]: + """ + 拉取历史数据，统计活跃段特征，自动推断预测策略和阈值。 + + 空闲段过滤：排除 p10 以下的点，避免机床空闲时的零值拉低阈值。 + strategy 判断：cv < 0.15 → phase_point（稳定信号），否则 phase_band（波动信号）。 + phase_lock 搜索范围：由周期长度的变异系数动态决定，周期抖动大则搜索范围宽。 + """ + ts_raw, ys_raw = fetch_history(f'{metric}{{device_id="{device_id}"}}') + if len(ys_raw) < MIN_POINTS: + return None + + arr = np.array(ys_raw, dtype=float) + + # 过滤空闲段：只保留活跃值（高于 p10） + p10_val = float(np.percentile(arr, 10)) + active = arr[arr > p10_val] + if len(active) < 30: + active = arr # 数据全是活跃段，不过滤 + + mean_val = float(np.mean(active)) + std_val = float(np.std(active)) + cv = std_val / max(abs(mean_val), 1e-6) + p5 = float(np.percentile(active, 5)) + p95 = float(np.percentile(active, 95)) + iqr = p95 - p5 + + # 策略自动判断 + strategy = "phase_point" if cv < 0.15 else "phase_band" + + # 阈值自动计算：取 IQR 的 80%、量程的 5%、2倍标准差 三者最大值 + abs_threshold = max(iqr * 0.8, (p95 - p5) * 0.05, std_val * 2.0) + rel_threshold = min(0.30, cv * 1.5) + + # phase_band 容忍带宽度：IQR 的 30% 或 1 倍标准差，取较大值 + band_pad_abs = max(iqr * 0.3, std_val) + + # phase-lock 搜索范围：从历史数据估算周期抖动率 + # 用 FFT 粗估周期，再用自相关精化，最后计算多周期长度的变异系数 + ts_grid, ys_grid = normalize_history(ts_raw, ys_raw) + period_search_ratio = PHASE_LOCK_PERIOD_SEARCH_RATIO # 默认值 + if len(ys_grid) >= MIN_POINTS: + rough_period = estimate_period_rough(ys_grid) + if rough_period > MIN_PERIOD_SECONDS: + # 用谷底间距估算周期抖动 + valleys = find_valley_indices(ts_grid, ys_grid, rough_period) + if len(valleys) >= 3: + diffs = np.diff(ts_grid[valleys].astype(float)) + valid = diffs[(diffs > rough_period * 0.5) & (diffs < rough_period * 2.0)] + if len(valid) >= 2: + period_cv = float(np.std(valid) / max(np.mean(valid), 1e-6)) + period_search_ratio = float(np.clip(period_cv * 2.0, 0.12, 0.25)) + + logger.info( + "推断指标特征 device=%s metric=%s cv=%.3f strategy=%s abs_thr=%.3f rel_thr=%.3f period_search=%.2f", + device_id, metric, cv, strategy, abs_threshold, rel_threshold, period_search_ratio, + ) + + return MetricProfile( + device_id=device_id, + metric=metric, + p5=p5, + p95=p95, + iqr=iqr, + cv=cv, + strategy=strategy, + abs_threshold=abs_threshold, + rel_threshold=rel_threshold, + band_low_q=5.0, + band_high_q=95.0, + band_pad_abs=band_pad_abs, + phase_lock_period_search_ratio=period_search_ratio, + ) + + +def load_overrides() -> Dict: + """加载人工上下限覆盖文件，文件不存在时返回空字典。""" + if not os.path.exists(OVERRIDE_FILE): + return {} + try: + with open(OVERRIDE_FILE, "r", encoding="utf-8") as f: + return json.load(f) + except Exception as e: + logger.warning("加载 override 文件失败 %s: %s", OVERRIDE_FILE, e) + return {} + + +def build_target(profile: MetricProfile, overrides: Dict) -> Dict: + """将 MetricProfile 转换为预测执行层可用的 target dict。""" + device_overrides = overrides.get(profile.device_id, {}).get(profile.metric, {}) + + target: Dict = { + "query": f'{profile.metric}{{device_id="{profile.device_id}"}}', + "pred_metric": f"{profile.metric}_predicted", + "anomaly_metric": f"{profile.metric}_anomaly", + "strategy": profile.strategy, + "abs_threshold": profile.abs_threshold, + "rel_threshold": profile.rel_threshold, + "smooth_window": 5 if profile.strategy == "phase_band" else 2, "outside_ratio_threshold": 0.60, "min_consecutive_outside": 5, "severe_exceed_ratio": 1.8, - }, - { - "query": 'vibration_x{device_id="fanuc-cnc"}', - "pred_metric": "vibration_x_predicted", - "anomaly_metric": "vibration_x_anomaly", - "strategy": "phase_band", - - # vibration 类指标噪声、尖峰较多，不建议用很窄的阈值。 - "abs_threshold": 0.18, - "rel_threshold": 0.55, - - # 平滑只用于相位锁定和 predicted 中位趋势。 - "smooth_window": 5, - - # upper/lower 用原始值分位数，范围放宽，覆盖正常尖峰。 - "band_low_q": 1, - "band_high_q": 99, - "band_pad_abs": 0.15, - - # 偶发越界容忍。 - "outside_ratio_threshold": 0.70, - "min_consecutive_outside": 5, - "severe_exceed_ratio": 2.0, - }, - { - "query": 'vibration_y{device_id="fanuc-cnc"}', - "pred_metric": "vibration_y_predicted", - "anomaly_metric": "vibration_y_anomaly", - "strategy": "phase_band", - "abs_threshold": 0.18, - "rel_threshold": 0.55, - "smooth_window": 5, - "band_low_q": 1, - "band_high_q": 99, - "band_pad_abs": 0.15, - "outside_ratio_threshold": 0.70, - "min_consecutive_outside": 5, - "severe_exceed_ratio": 2.0, - }, - { - "query": 'vibration_z{device_id="fanuc-cnc"}', - "pred_metric": "vibration_z_predicted", - "anomaly_metric": "vibration_z_anomaly", - "strategy": "phase_band", - "abs_threshold": 0.18, - "rel_threshold": 0.55, - "smooth_window": 5, - "band_low_q": 1, - "band_high_q": 99, - "band_pad_abs": 0.15, - "outside_ratio_threshold": 0.70, - "min_consecutive_outside": 5, - "severe_exceed_ratio": 2.0, - }, -] + "phase_lock_period_search_ratio": profile.phase_lock_period_search_ratio, + "phase_lock_origin_search_ratio": min( + 0.45, profile.phase_lock_period_search_ratio * 2.5 + ), + # 物理上下限（可选，来自 override 文件） + "hard_max": device_overrides.get("hard_max"), + "hard_min": device_overrides.get("hard_min"), + } + + if profile.strategy == "phase_band": + target.update({ + "band_low_q": profile.band_low_q, + "band_high_q": profile.band_high_q, + "band_pad_abs": profile.band_pad_abs, + }) + + return target + + +def refresh_targets_if_needed() -> None: + """ + 按 TARGETS_REFRESH_INTERVAL 间隔重新发现设备和指标，动态更新目标列表。 + 首次调用时立即执行发现。 + """ + global _TARGETS_CACHE, _TARGETS_LAST_REFRESH + + now = time.time() + if now - _TARGETS_LAST_REFRESH < TARGETS_REFRESH_INTERVAL and _TARGETS_CACHE: + return + + logger.info("开始发现设备和指标...") + overrides = load_overrides() + targets: List[Dict] = [] + + device_ids = discover_device_ids() + if not device_ids: + logger.warning("未发现任何 device_id，保持现有目标列表") + return + + for device_id in device_ids: + metrics = discover_metrics_for_device(device_id) + for metric in metrics: + profile = infer_metric_profile(device_id, metric) + if profile is not None: + targets.append(build_target(profile, overrides)) + + if targets: + _TARGETS_CACHE = targets + _TARGETS_LAST_REFRESH = now + logger.info( + "目标列表已更新：%d 台设备，%d 个指标目标", + len(device_ids), + len(targets), + ) + else: + logger.warning("发现流程未产生任何有效目标，保持现有目标列表") -EXTRA_PREDICT_LABELS = { - "forecast": "phase_band_health_v12", - "source": "protoforge", -} BASELINE_STATUS_HEALTHY = "healthy" BASELINE_STATUS_ANOMALY = "anomaly" @@ -612,8 +772,6 @@ def build_templates_from_valleys( if strategy == "phase_band": mid_template = np.percentile(mid_arr, 50, axis=0) - - # upper/lower 使用原始值分布，而不是平滑值分布。 lower_template = np.percentile(band_arr, low_q, axis=0) upper_template = np.percentile(band_arr, high_q, axis=0) else: @@ -794,17 +952,28 @@ def merge_template( # ============================================================================= # Phase Lock +# 支持 target 级别的 phase_lock_period_search_ratio / phase_lock_origin_search_ratio +# 粗铣工位周期含随机抖动(±10s)，需要更宽的搜索范围 # ============================================================================= def phase_lock_recent( state: BaselineState, ts_grid: np.ndarray, ys_model: np.ndarray, + target: Optional[Dict] = None, ) -> Tuple[int, int, np.ndarray, float]: base_period = int(state.period) base_origin = int(state.phase_origin_ts) base_template = np.array(state.template, dtype=float) + # 从 target 读取搜索范围，允许粗铣工位使用更宽的范围 + period_search_ratio = float( + (target or {}).get("phase_lock_period_search_ratio", PHASE_LOCK_PERIOD_SEARCH_RATIO) + ) + origin_search_ratio = float( + (target or {}).get("phase_lock_origin_search_ratio", PHASE_LOCK_ORIGIN_SEARCH_RATIO) + ) + if base_period <= 1 or len(base_template) <= 1: ts_recent = ts_grid[-DETECT_WINDOW_SECONDS:].astype(int).tolist() pred = predict_template_values(base_template, base_period, base_origin, ts_recent) @@ -832,11 +1001,11 @@ def phase_lock_recent( p_min = max( int(MIN_PERIOD_SECONDS), - int(round(base_period * (1.0 - PHASE_LOCK_PERIOD_SEARCH_RATIO))), + int(round(base_period * (1.0 - period_search_ratio))), ) p_max = min( int(MAX_PERIOD_SECONDS), - int(round(base_period * (1.0 + PHASE_LOCK_PERIOD_SEARCH_RATIO))), + int(round(base_period * (1.0 + period_search_ratio))), ) best_period = base_period @@ -855,7 +1024,7 @@ def phase_lock_recent( for period in range(p_min, p_max + 1, PHASE_LOCK_PERIOD_STEP): template = resample_template(base_template, period) center_origin = normalize_origin_near(base_origin, period, last_ts) - origin_shift = max(2, int(round(period * PHASE_LOCK_ORIGIN_SEARCH_RATIO))) + origin_shift = max(2, int(round(period * origin_search_ratio))) for shift in range(-origin_shift, origin_shift + 1, PHASE_LOCK_ORIGIN_STEP): origin = center_origin + shift @@ -925,7 +1094,6 @@ def calc_final_bounds( if strategy == "phase_band": pad_abs = float(target.get("band_pad_abs", abs_threshold)) - # 对 vibration 类指标：边界更像正常波动容忍带，不是硬边界。 dynamic_pad = np.maximum( pad_abs, np.abs(pred) * rel_threshold * 0.25, @@ -933,10 +1101,18 @@ def calc_final_bounds( lower = lower_raw - dynamic_pad upper = upper_raw + dynamic_pad + else: + lower, upper = calc_point_bounds(pred, abs_threshold, rel_threshold) - return lower, upper + # 物理上下限兜底（来自 override 文件，可选） + hard_max = target.get("hard_max") + hard_min = target.get("hard_min") + if hard_max is not None: + upper = np.minimum(upper, float(hard_max)) + if hard_min is not None: + lower = np.maximum(lower, float(hard_min)) - return calc_point_bounds(pred, abs_threshold, rel_threshold) + return lower, upper def detect_anomaly( @@ -950,6 +1126,7 @@ def detect_anomaly( state=state, ts_grid=ts_grid, ys_model=ys_model, + target=target, ) recent_len = len(pred_recent) @@ -1018,11 +1195,6 @@ def detect_anomaly( target.get("severe_exceed_ratio", SEVERE_EXCEED_RATIO) ) - # 核心优化： - # 1. 偶发 1~3 个点越界不报警。 - # 2. 持续越界才报警。 - # 3. 高比例越界才报警。 - # 4. 严重越界才立即报警。 is_anomaly = ( outside_ratio >= outside_ratio_threshold or max_outside_seconds >= min_consecutive_outside @@ -1653,7 +1825,13 @@ def build_prediction_timestamps( def run_once() -> None: now_str = datetime.now().strftime("%H:%M:%S") - for target in PREDICT_TARGETS: + refresh_targets_if_needed() + + if not _TARGETS_CACHE: + logger.warning("[%s] 目标列表为空，等待设备发现完成", now_str) + return + + for target in _TARGETS_CACHE: query = target["query"] pred_metric = target["pred_metric"] anomaly_metric = target["anomaly_metric"] @@ -1749,7 +1927,7 @@ def run_once() -> None: origin_str = datetime.fromtimestamp(state.phase_origin_ts).strftime("%H:%M:%S") logger.info( - "[%s] %-40s → %-35s strategy=%s status=%s anomaly=%s outside=%.2f max_outside=%ss max_exceed=%.2f period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", + "[%s] %-50s → %-35s strategy=%s status=%s anomaly=%s outside=%.2f max_outside=%ss max_exceed=%.2f period=%ss origin=%s last_real=%s lag=%ss 写入 %d 点，预测区间 %s ~ %s", now_str, query, pred_metric, @@ -1775,7 +1953,7 @@ def main() -> None: load_state() logger.info( - "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s", + "预测服务启动 VM=%s 历史窗口=%dmin 理论预测窗口=%ds 实际写入窗口=%ds 轮询间隔=%ds state=%s forecast=%s override=%s refresh=%ds", VM_URL, HISTORY_MINUTES, HORIZON_SECONDS, @@ -1783,6 +1961,8 @@ def main() -> None: POLL_INTERVAL, STATE_FILE, EXTRA_PREDICT_LABELS["forecast"], + OVERRIDE_FILE, + TARGETS_REFRESH_INTERVAL, ) while True: @@ -1791,4 +1971,4 @@ def main() -> None: if __name__ == "__main__": - main() \ No newline at end of file + main() diff --git a/protoforge/core/fault.py b/protoforge/core/fault.py index 11b61a7..2182bf1 100644 --- a/protoforge/core/fault.py +++ b/protoforge/core/fault.py @@ -32,24 +32,77 @@ BUILTIN_FAULT_TYPES: list[FaultTypeDefinition] = [ # ------------------------------------------------------------------ - # 进给堵转 — 工件夹紧松动或切削量过大导致进给卡死 - # 特征：进给速率瞬间降为0，主轴负载和电流急剧升高，主轴仍在转（区别于崩刃） - # 模式：瞬间注入 + # 进给堵转（粗铣）— fanuc-cnc + # 量程：spindle_speed~2000RPM, feed_rate~800mm/min, + # spindle_current~21A, spindle_load~56% + # 堵转目标：load→92%, current→38A，转速维持+轻微抖动 # ------------------------------------------------------------------ FaultTypeDefinition( - id="feed_stall", - name="进给堵转", - description="进给轴卡死，进给速率降为零，主轴负载和电流急剧升高，主轴转速维持（区别于崩刃停主轴）", + id="feed_stall_rough", + name="进给堵转（粗铣）", + description="粗铣进给轴卡死，进给速率降为零，主轴负载升至~92%，电流升至~38A，主轴转速维持（区别于崩刃停主轴）", category="process", default_duration=20.0, - tags=["进给", "堵转", "突发"], + tags=["进给", "堵转", "突发", "粗铣"], point_faults=[ PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, - multiplier=2.8, noise_scale=5.0), + target_value=92.0, noise_scale=4.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - multiplier=3.8, noise_scale=1.5), + target_value=38.0, noise_scale=1.5), + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=30.0), + ], + ), + + # ------------------------------------------------------------------ + # 进给堵转（半精铣）— fanuc-cnc-semi-finish + # 量程：spindle_speed~4000RPM, feed_rate~500mm/min, + # spindle_current~14.5A, spindle_load~38% + # 堵转目标：load→68%, current→26A，转速维持+轻微抖动 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="feed_stall_semi", + name="进给堵转（半精铣）", + description="半精铣进给轴卡死，进给速率降为零，主轴负载升至~68%，电流升至~26A，主轴转速维持（区别于崩刃停主轴）", + category="process", + default_duration=20.0, + tags=["进给", "堵转", "突发", "半精铣"], + point_faults=[ + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + target_value=0.0, noise_scale=0.0), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + target_value=68.0, noise_scale=3.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + target_value=26.0, noise_scale=1.2), + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=50.0), + ], + ), + + # ------------------------------------------------------------------ + # 进给堵转（精铣）— fanuc-cnc-finish + # 量程：spindle_speed~6000RPM, feed_rate~300mm/min, + # spindle_current~8.5A, spindle_load~22% + # 堵转目标：load→40%, current→15A，转速维持+轻微抖动 + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="feed_stall_finish", + name="进给堵转（精铣）", + description="精铣进给轴卡死，进给速率降为零，主轴负载升至~40%，电流升至~15A，主轴转速维持（区别于崩刃停主轴）", + category="process", + default_duration=20.0, + tags=["进给", "堵转", "突发", "精铣"], + point_faults=[ + PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, + target_value=0.0, noise_scale=0.0), + PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, + target_value=40.0, noise_scale=2.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, + target_value=15.0, noise_scale=0.8), + PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, + multiplier=1.0, noise_scale=80.0), ], ), From 07fc5d6897f2ab7b228fa16d748cd9ad1c74abb9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=BC=A0=E5=B0=91=E6=99=BA?= Date: Fri, 29 May 2026 10:09:13 +0800 Subject: [PATCH 36/36] fix(fault): fault fix --- FAULT_INJECTION.md | 47 ++++++++++--- protoforge/core/fault.py | 133 +++++++++++++++++++++++++++---------- protoforge/models/fault.py | 8 ++- 3 files changed, 142 insertions(+), 46 deletions(-) diff --git a/FAULT_INJECTION.md b/FAULT_INJECTION.md index 951648d..22746bf 100644 --- a/FAULT_INJECTION.md +++ b/FAULT_INJECTION.md @@ -160,19 +160,48 @@ DELETE /api/v1/devices/{device_id}/fault --- -### spindle_overheat — 主轴过热 +### spindle_overheat_rough — 主轴过热（粗铣） - **分类**：thermal -- **模式**：渐进式 +- **模式**：渐进式（绝对目标值） - **默认持续时间**：240 秒 -- **真实场景**：长时间高负荷或冷却系统故障，热保护机制逐渐降低转速 +- **真实场景**：粗铣主轴长时间高负荷或冷却不足，负载/电流持续高位，热保护渐进降速 -| 测点 | 变化方向 | 峰值倍率 | -|------|---------|---------| -| `spindle_current` | 升高 | ×1.8 | -| `spindle_speed` | 降低 | ×0.6 | -| `vibration_x` | 升高 | ×1.5 | -| `vibration_z` | 升高 | ×1.5 | +| 测点 | 变化方向 | 目标值 | +|------|---------|--------| +| `spindle_load` | 持续升高 | →85% | +| `spindle_current` | 持续升高 | →34A | +| `spindle_speed` | 渐进降低 | →1400 RPM | + +--- + +### spindle_overheat_semi — 主轴过热（半精铣） + +- **分类**：thermal +- **模式**：渐进式（绝对目标值） +- **默认持续时间**：240 秒 +- **真实场景**：半精铣主轴长时间高负荷或冷却不足，负载/电流持续高位，热保护渐进降速 + +| 测点 | 变化方向 | 目标值 | +|------|---------|--------| +| `spindle_load` | 持续升高 | →72% | +| `spindle_current` | 持续升高 | →24A | +| `spindle_speed` | 渐进降低 | →2600 RPM | + +--- + +### spindle_overheat_finish — 主轴过热（精铣） + +- **分类**：thermal +- **模式**：渐进式（绝对目标值） +- **默认持续时间**：240 秒 +- **真实场景**：精铣主轴长时间高负荷或冷却不足，负载/电流持续高位，热保护渐进降速 + +| 测点 | 变化方向 | 目标值 | +|------|---------|--------| +| `spindle_load` | 持续升高 | →48% | +| `spindle_current` | 持续升高 | →15A | +| `spindle_speed` | 渐进降低 | →3800 RPM | --- diff --git a/protoforge/core/fault.py b/protoforge/core/fault.py index 2182bf1..857bf90 100644 --- a/protoforge/core/fault.py +++ b/protoforge/core/fault.py @@ -35,12 +35,12 @@ # 进给堵转（粗铣）— fanuc-cnc # 量程：spindle_speed~2000RPM, feed_rate~800mm/min, # spindle_current~21A, spindle_load~56% - # 堵转目标：load→92%, current→38A，转速维持+轻微抖动 + # 堵转目标：load→85~100%, current→34~42A，转速维持+轻微抖动 # ------------------------------------------------------------------ FaultTypeDefinition( id="feed_stall_rough", name="进给堵转（粗铣）", - description="粗铣进给轴卡死，进给速率降为零，主轴负载升至~92%，电流升至~38A，主轴转速维持（区别于崩刃停主轴）", + description="粗铣进给轴卡死，进给速率降为零，主轴负载升至85~100%，电流升至34~42A，主轴转速维持（区别于崩刃停主轴）", category="process", default_duration=20.0, tags=["进给", "堵转", "突发", "粗铣"], @@ -48,9 +48,9 @@ PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, - target_value=92.0, noise_scale=4.0), + target_min=85.0, target_max=100.0, noise_scale=4.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - target_value=38.0, noise_scale=1.5), + target_min=34.0, target_max=42.0, noise_scale=1.5), PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, multiplier=1.0, noise_scale=30.0), ], @@ -60,12 +60,12 @@ # 进给堵转（半精铣）— fanuc-cnc-semi-finish # 量程：spindle_speed~4000RPM, feed_rate~500mm/min, # spindle_current~14.5A, spindle_load~38% - # 堵转目标：load→68%, current→26A，转速维持+轻微抖动 + # 堵转目标：load→62~75%, current→23~29A，转速维持+轻微抖动 # ------------------------------------------------------------------ FaultTypeDefinition( id="feed_stall_semi", name="进给堵转（半精铣）", - description="半精铣进给轴卡死，进给速率降为零，主轴负载升至~68%，电流升至~26A，主轴转速维持（区别于崩刃停主轴）", + description="半精铣进给轴卡死，进给速率降为零，主轴负载升至62~75%，电流升至23~29A，主轴转速维持（区别于崩刃停主轴）", category="process", default_duration=20.0, tags=["进给", "堵转", "突发", "半精铣"], @@ -73,9 +73,9 @@ PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, - target_value=68.0, noise_scale=3.0), + target_min=62.0, target_max=75.0, noise_scale=3.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - target_value=26.0, noise_scale=1.2), + target_min=23.0, target_max=29.0, noise_scale=1.2), PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, multiplier=1.0, noise_scale=50.0), ], @@ -85,12 +85,12 @@ # 进给堵转（精铣）— fanuc-cnc-finish # 量程：spindle_speed~6000RPM, feed_rate~300mm/min, # spindle_current~8.5A, spindle_load~22% - # 堵转目标：load→40%, current→15A，转速维持+轻微抖动 + # 堵转目标：load→36~45%, current→13~17A，转速维持+轻微抖动 # ------------------------------------------------------------------ FaultTypeDefinition( id="feed_stall_finish", name="进给堵转（精铣）", - description="精铣进给轴卡死，进给速率降为零，主轴负载升至~40%，电流升至~15A，主轴转速维持（区别于崩刃停主轴）", + description="精铣进给轴卡死，进给速率降为零，主轴负载升至36~45%，电流升至13~17A，主轴转速维持（区别于崩刃停主轴）", category="process", default_duration=20.0, tags=["进给", "堵转", "突发", "精铣"], @@ -98,33 +98,79 @@ PointFaultConfig(point="feed_rate", mode=FaultMode.INSTANT, target_value=0.0, noise_scale=0.0), PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, - target_value=40.0, noise_scale=2.0), + target_min=36.0, target_max=45.0, noise_scale=2.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - target_value=15.0, noise_scale=0.8), + target_min=13.0, target_max=17.0, noise_scale=0.8), PointFaultConfig(point="spindle_speed", mode=FaultMode.INSTANT, multiplier=1.0, noise_scale=80.0), ], ), # ------------------------------------------------------------------ - # 主轴过热 — 长时间高负荷或冷却系统故障 - # 特征：主轴负载和电流持续偏高，转速因热保护逐渐降低 - # 模式：渐进式，持续时间较长 + # 主轴过热（粗铣）— fanuc-cnc + # 基线：spindle_speed~2000RPM, spindle_current~21A, spindle_load~56% + # 过热目标范围：load 78~92%，current 30~38A，转速降至 1200~1600RPM + # 范围模拟不同冷却状态、负荷历史、环境温度下的个体差异 + # 模式：渐进式；全部用 target_min/max，避免 multiplier 在空载基线=0 时失效 # ------------------------------------------------------------------ FaultTypeDefinition( - id="spindle_overheat", - name="主轴过热", - description="主轴长时间高负荷运转或冷却不足，spindle_load和spindle_current持续偏高，转速因热保护渐进下降", + id="spindle_overheat_rough", + name="主轴过热（粗铣）", + description="粗铣主轴长时间高负荷或冷却不足，spindle_load渐进升至78~92%，spindle_current升至30~38A，转速因热保护渐进降至1200~1600RPM", category="thermal", default_duration=240.0, - tags=["主轴", "过热", "渐进"], + tags=["主轴", "过热", "渐进", "粗铣"], point_faults=[ PointFaultConfig(point="spindle_load", mode=FaultMode.GRADUAL, - multiplier=1.6, noise_scale=3.0), + target_min=78.0, target_max=92.0, noise_scale=3.5), PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, - multiplier=1.8, noise_scale=1.2), + target_min=30.0, target_max=38.0, noise_scale=1.5), PointFaultConfig(point="spindle_speed", mode=FaultMode.GRADUAL, - multiplier=0.6, noise_scale=50.0), + target_min=1200, target_max=1600, noise_scale=40.0), + ], + ), + + # ------------------------------------------------------------------ + # 主轴过热（半精铣）— fanuc-cnc-semi-finish + # 基线：spindle_speed~4000RPM, spindle_current~14.5A, spindle_load~38% + # 过热目标范围：load 65~78%，current 21~27A，转速降至 2400~2900RPM + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="spindle_overheat_semi", + name="主轴过热（半精铣）", + description="半精铣主轴长时间高负荷或冷却不足，spindle_load渐进升至65~78%，spindle_current升至21~27A，转速因热保护渐进降至2400~2900RPM", + category="thermal", + default_duration=240.0, + tags=["主轴", "过热", "渐进", "半精铣"], + point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.GRADUAL, + target_min=65.0, target_max=78.0, noise_scale=3.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, + target_min=21.0, target_max=27.0, noise_scale=1.2), + PointFaultConfig(point="spindle_speed", mode=FaultMode.GRADUAL, + target_min=2400, target_max=2900, noise_scale=50.0), + ], + ), + + # ------------------------------------------------------------------ + # 主轴过热（精铣）— fanuc-cnc-finish + # 基线：spindle_speed~6000RPM, spindle_current~8.5A, spindle_load~22% + # 过热目标范围：load 42~55%，current 13~17A，转速降至 3600~4200RPM + # ------------------------------------------------------------------ + FaultTypeDefinition( + id="spindle_overheat_finish", + name="主轴过热（精铣）", + description="精铣主轴长时间高负荷或冷却不足，spindle_load渐进升至42~55%，spindle_current升至13~17A，转速因热保护渐进降至3600~4200RPM", + category="thermal", + default_duration=240.0, + tags=["主轴", "过热", "渐进", "精铣"], + point_faults=[ + PointFaultConfig(point="spindle_load", mode=FaultMode.GRADUAL, + target_min=42.0, target_max=55.0, noise_scale=2.0), + PointFaultConfig(point="spindle_current", mode=FaultMode.GRADUAL, + target_min=13.0, target_max=17.0, noise_scale=0.8), + PointFaultConfig(point="spindle_speed", mode=FaultMode.GRADUAL, + target_min=3600, target_max=4200, noise_scale=60.0), ], ), @@ -230,15 +276,15 @@ FaultTypeDefinition( id="air_cutting", name="空切检测", - description="刀具未接触工件，spindle_load跌至空载区间(5-15%)，spindle_current降至空转水平，转速进给保持正常", + description="刀具未接触工件，spindle_load跌至空载区间(4-12%)，spindle_current降至空转水平，转速进给保持正常", category="tool", default_duration=180.0, tags=["刀具", "空切", "工况切换", "负载"], point_faults=[ PointFaultConfig(point="spindle_load", mode=FaultMode.INSTANT, - target_value=8.0, noise_scale=2.0), + target_min=4.0, target_max=12.0, noise_scale=2.0), PointFaultConfig(point="spindle_current", mode=FaultMode.INSTANT, - target_value=2.5, noise_scale=0.3), + target_min=2.0, target_max=3.5, noise_scale=0.3), ], ), @@ -356,6 +402,13 @@ def inject(self, device: Any, request: FaultInjectRequest) -> FaultInfo: except (TypeError, ValueError): baseline[pf.point] = 0.0 + # 对有范围定义的测点，注入时随机采样一个实际目标值 + # 使每次注入的故障严重程度有所不同，模拟真实场景的个体差异 + resolved_targets: dict[str, float] = {} + for pf in fault_type.point_faults: + if pf.target_min is not None and pf.target_max is not None: + resolved_targets[pf.point] = random.uniform(pf.target_min, pf.target_max) + fault = ActiveFault( fault_id=uuid.uuid4().hex[:12], device_id=device.id, @@ -365,10 +418,11 @@ def inject(self, device: Any, request: FaultInjectRequest) -> FaultInfo: duration=duration, started_at=time.time(), baseline_values=baseline, + resolved_targets=resolved_targets, ) self._active[device.id] = fault - logger.info("Fault injected: device=%s type=%s duration=%.0fs", - device.id, fault_type.id, duration) + logger.info("Fault injected: device=%s type=%s duration=%.0fs resolved_targets=%s", + device.id, fault_type.id, duration, resolved_targets) return self._to_info(fault, fault_type) def apply(self, device: Any) -> None: @@ -400,13 +454,14 @@ def apply(self, device: Any) -> None: baseline = fault.baseline_values.get(pf.point, 0.0) if baseline == 0.0: # 基线为0说明注入时设备处于换刀/停机状态 - # target_value 模式可以直接执行（如崩刃归零、空切归空载） + # target_value / resolved_targets 模式可以直接执行 # multiplier 模式跳过，避免在零基线上产生无意义的值 - if pf.target_value is None: + if pf.target_value is None and pf.point not in fault.resolved_targets: continue + resolved_target = fault.resolved_targets.get(pf.point) device._point_values[pf.point] = self._compute_value( - pf, baseline, progress, fault.intensity + pf, baseline, progress, fault.intensity, resolved_target ) def clear(self, device_id: str) -> bool: @@ -451,20 +506,26 @@ def _compute_value( baseline: float, progress: float, intensity: float, + resolved_target: Optional[float] = None, ) -> float: - """根据故障配置和当前进度计算覆盖值""" + """根据故障配置和当前进度计算覆盖值。 + + 目标值优先级：resolved_target（注入时随机采样）> target_value（固定值）> multiplier + """ + # 确定本次注入的实际目标值 + effective_target: Optional[float] = resolved_target if resolved_target is not None else pf.target_value + if pf.mode == FaultMode.INSTANT: - # 瞬间模式：直接用目标值，不随时间变化 - if pf.target_value is not None: - target = pf.target_value + if effective_target is not None: + target = effective_target elif pf.multiplier is not None: target = baseline * (1.0 + (pf.multiplier - 1.0) * intensity) else: target = baseline else: # 渐进模式：随 progress 线性劣化 - if pf.target_value is not None: - target = baseline + (pf.target_value - baseline) * progress * intensity + if effective_target is not None: + target = baseline + (effective_target - baseline) * progress * intensity elif pf.multiplier is not None: target = baseline * (1.0 + (pf.multiplier - 1.0) * progress * intensity) else: diff --git a/protoforge/models/fault.py b/protoforge/models/fault.py index cc038e0..9928332 100644 --- a/protoforge/models/fault.py +++ b/protoforge/models/fault.py @@ -25,6 +25,11 @@ class PointFaultConfig(BaseModel): target_value: Optional[float] = None multiplier: Optional[float] = None # 异常值 = 当前正常值 × multiplier + # 目标值范围：注入时在 [target_min, target_max] 内随机采样一个实际目标值 + # 设置后会覆盖 target_value，使每次注入的故障严重程度有所不同 + target_min: Optional[float] = None + target_max: Optional[float] = None + # GRADUAL 模式：从当前值线性劣化到 target_value 或 multiplier 倍 # 劣化程度 = progress(0~1) × (target - baseline) noise_scale: float = 0.0 # 叠加随机噪声幅度，模拟真实抖动 @@ -59,7 +64,8 @@ class ActiveFault(BaseModel): duration: float = 120.0 started_at: float = 0.0 cleared_at: Optional[float] = None - baseline_values: dict[str, float] = Field(default_factory=dict) # 注入时的正常基线值 + baseline_values: dict[str, float] = Field(default_factory=dict) # 注入时的正常基线值 + resolved_targets: dict[str, float] = Field(default_factory=dict) # 注入时随机采样的实际目标值 class FaultInfo(BaseModel):