TinyML for Anomaly Detection

Tiny machine learning (TinyML) is revolutionizing anomaly detection by enabling the deployment of intelligent models directly on edge devices, especially within Internet of Things (IoT) systems. This chapter explores the integration of TinyML in real-time monitoring systems, where low-latency responses and reduced communication overheads are crucial. TinyML offers key advantages, including energy efficiency, enhanced privacy, and the ability to operate with minimal reliance on cloud resources. It enables anomaly detection in time-series and image data using machine learning models that can be adapted to constrained environments through techniques such as model quantization and one-class learning. Despite its potential, TinyML faces several challenges, including limited memory and processing power, as well as data availability and preprocessing constraints on the edge. In the context of edge computing (EC), which allows for processing directly on IoT devices, TinyML plays a pivotal role in Artificial Intelligence of Things (AIoT). This field focuses on implementing machine and deep learning models on microcontroller units (MCUs), facilitating anomaly detection at the data source. A systematic literature review of 20 research papers from 2021 to 2024 sheds light on various anomaly detection techniques and their applications in EC. The review encompasses essential aspects, including the types of machine learning (ML) and deep learning (DL) models employed, data sources for model training, microcontroller hardware, power supply methods, and communication technologies. The study also develops a taxonomy of ML/DL algorithms used for anomaly detection in TinyML and examines the benefits and challenges faced in applying these techniques to constrained devices. This comprehensive analysis offers valuable insights into the expanding role of TinyML in deploying scalable, real-time anomaly detection systems across various IoT applications.