Self-Supervised Online and Lightweight Anomaly and Event Detection for IoT Devices

Abstract

The increasing number of Internet of Things (IoT) devices and low-cost sensors have facilitated developments in large-scale monitoring applications. However, the accuracy of low-cost sensors remains questionable. Monitoring applications, such as environmental monitoring, try to detect “interesting” data points or patterns, known as anomalies, that do not conform to the norm. These include erroneous data caused by hardware failures or malicious attacks, and nonerroneous data due to unexpected phenomenon, caused by events, such as unexpected high traffic volume. Traditionally, IoT devices collect raw data and periodically upload them to the cloud for processing, which includes anomaly detection. However, the increasing processing capabilities of IoT devices have made the on-device anomaly detection possible in an online and real-time manner. In this article, multivariate long short-term memory (LSTM) autoencoder is proposed for anomaly and event detection in IoT devices. In addition, the proposed approach integrates smart inference, based on a game-theoretical approach, which dynamically changes the period of detection based on the stability of the data, aiming to optimize power consumption and elongate the lifetime of the device. The proposed anomaly and event detection model was simulated and implemented on an STM32H743 Nucleo board, and results show the robustness of the model regardless of the number of anomalies and events present.