An Adaptable and Unsupervised TinyML Anomaly Detection System for Extreme Industrial Environments

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Industrial assets rely on sensors for status monitoring, with data analysis typically performed remotely via cloud or data centers.
• Remote analysis presents challenges including system complexity, connectivity dependence, and unsuitability for extreme environments.
• Existing methods are limited by communication restrictions and potential failures in isolated or harsh industrial settings.

Purpose of the Study:

• To propose and evaluate an end-to-end adaptable anomaly detection system for extreme industrial environments.
• To leverage Internet of Things (IoT), edge computing, and Tiny Machine Learning (TinyML) for real-time anomaly detection.
• To address limitations of cloud-based analysis in terms of connectivity and operational constraints.

Main Methods:

• Development of an IoT sensing Kit with an ESP32 microcontroller and MicroPython firmware for edge data processing.
• Implementation of an anomaly detection model using the isolation forest algorithm trainable on the microcontroller.
• Integration of blockchain technology for a secure and transparent anomaly record.

Main Results:

• The isolation forest model trains on the microcontroller in 1.2 to 6.4 seconds.
• Anomaly detection occurs in under 16 milliseconds using 50 trees and 80 KB of RAM.
• The system successfully operates in an extreme industrial environment, demonstrating adaptability and configurability.

Conclusions:

• The proposed edge computing system effectively performs real-time anomaly detection in extreme industrial settings.
• TinyML and IoT integration on edge devices offer a robust and efficient alternative to cloud-based analysis.
• The system enhances industrial asset monitoring by providing timely failure alerts and a secure anomaly log.