Time Series Anomaly Detection in Vehicle Sensors Using Self-Attention Mechanisms

Abstract

Connected autonomous vehicles (CAVs) offer significant enhancements in coordinated traffic and safety through real-time vehicle-to-vehicle or vehicle-to-infrastructure communications, establishing them as a potent tool for augmenting driving tasks. However, the extensive information-sharing framework inherent in CAVs amplifies the risk associated with sensor anomalies, posing challenges to the reliability and security of the system. Responding to this timely research challenge, this study proposes a novel anomaly detection method, namely Dual-channel Self-attention-based Convolutional Neural Network (DSA-CNN) for multivariate time series data. Through the introduction of the Dual-channel Self-attention Mechanism, DSA-CNN can progressively and autonomously extract spatiotemporal features from multivariate time series data. The proposed method was tested under a variety of common threatening sensor anomaly patterns of CAVs summarised in the literature, and evaluated under multiple different performance metrics. The results demonstrate its advantages in detecting minor anomalies and enhancing sensitivity, outperforming previously reported methods in the literature. Across all experimental scenarios, an average sensitivity improvement of 2.53% was observed, complemented by an average F1 score increase of 1.47%. In CAV settings, maintaining high sensitivity to ensure fewer undetected anomalies, alongside the ability to detect small anomalies, can be more important for the robustness and safety measures of CAV systems.