On-Device Training Based Anomaly Detection Platform

Bao-Toan Thai; Vy-Khang Tran; Chi-Ngon Nguyen; Van-Khanh Nguyen

Authors

Bao-Toan Thai PLC Technology and IIoT Laboratory, College of Engineering, Can Tho University https://orcid.org/0009-0005-7036-1893
Vy-Khang Tran PLC Technology and IIoT Laboratory, College of Engineering, Can Tho University https://orcid.org/0000-0002-4321-0126
Chi-Ngon Nguyen College of Engineering, Can Tho University https://orcid.org/0000-0002-9638-7259
Van-Khanh Nguyen Can Tho University https://orcid.org/0000-0003-0006-6803

Keywords:

On-device training, autoencoder, anomaly detection, microcontroller unit, tiny machine learning (tinyML)

Abstract

Anomaly Detection (AD) identifies issues in monitored systems, allowing timely responses. It's used in areas like industrial systems, healthcare, and networks. Among various AD methods, Autoencoder (AE)-based models are popular for their unsupervised training, which doesn’t require labeled data. This study proposes a full platform that implements AE-based AD on tiny devices and tests it on motor operating noise. The hardware includes two microcontroller units (MCUs) responsible for (1) real-time data acquisition and processing, (2) AE training and threshold setting, and (3) running inference for anomaly alerts. To boost training speed, the platform supports both Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (Adam). The algorithm uses real-time data instead of pre-collected data to improve practical performance. Tests show the platform effectively trains on normal motor noise and accurately sets the anomaly threshold. It achieves a real-time anomaly detection accuracy of 100%. The system runs fully automatically, making it suitable for integration into edge AIoT systems.