On-Device Tiny Machine Learning for Anomaly Detection Based on the Extreme Values Theory

The significance of anomaly detection is particularly pronounced in Industry 4.0 applications. For instance, in manufacturing, the timely detection of equipment malfunctions can prevent costly downtime and maintain production efficiency. In energy systems, spotting anomalies in power consumption patterns can enhance resource allocation and optimize energy usage. Equally noteworthy is the ascendancy of tiny machine learning (TinyML), emerging as a potent tool for real-time anomaly detection, exemplifying its versatile utility. This work presents an unsupervised on-device learning TinyML algorithm, drawing inspiration from the extreme value theory. The algorithm leverages the two-parameter Weibull distribution function to efficiently identify anomalies within discrete time series data. Optimal hyperparameters are ascertained via grid search methodology. Notably, employing synthetic data with randomized anomalies elucidates the algorithm's proficiency in binary classification within time series, highlighting an accuracy of 99.80%, recall of 93.10%, and F1 score of 96.43%. The amalgamation of theoretical foundations from the extreme value theory and practical capabilities of TinyML accentuates its pertinence across a broad spectrum of domains.