TinyML-based anomaly detection system running on the AK Embedded Base Kit (STM32L151 + ICM-20948 IMU).
3-axis accelerometer data is collected at 58 Hz, processed through a DSP pipeline (Butterworth lowpass, FFT, PSD), and classified by a small fully-connected neural network to detect 4 motion states: idle, left-right, up-down, and vigorous shaking (maritine).
The entire pipeline from sensor sampling to feature extraction to neural inference and runs on the Cortex-M3 using CMSIS-DSP, with no cloud or external compute required.
Anomal Detection documentation: application/sources/app/nn/docs/Readme.md
| Path | Description |
|---|---|
| Trainning | Colab notebook, dataset, scaler computation script |
| AI Pipeline | On device inference engine (C/C++ + CMSIS-DSP) |
| Model | Model weights |
| Sensor | ICM-20948 handle |
- MCU: STM32L151 (ARM Cortex-M3)
- IMU: ICM-20948 (9-DoF, I2C)
| Topic | Link |
|---|---|
| Tutorials | https://epcb.vn/blogs/ak-embedded-software |
| Vendor | https://epcb.vn/products/ak-embedded-base-kit-lap-trinh-nhung-vi-dieu-khien-mcu |