Multivariate Time-Series Anomaly Detection with Temporal Self-Supervision and Graphs: Application to Vehicle Failure Prediction

Abstract

Failure prediction is key to ensuring the reliable operation of vehicles, especially for organizations that depend on a fleet of vehicles. However, traditional approaches often rely on rule-based or heuristic methods that may not be effective in detecting subtle anomalies, rare events, or in more modern vehicles containing a complex sensory network. This paper presents a novel approach to vehicle failure prediction, called mVSG-VFP, which employs self-supervised learning and graph-based techniques. The proposed method realizes the failure prediction task by exploring information hidden in the time-series data recorded through the sensors embedded in the vehicle. mVSG-VFP includes two main components: a graph-based autoencoder that learns representations of normal data while considering the relationship between different sensors and a self-supervised component that maps temporally-adjacent data to similar representations. We propose a novel approach to define the notion of adjacency in vehicle temporal data. To evaluate mVSG-VFP, we apply it to a dataset comprised of vehicle sensor recordings to identify the abnormal data samples that signal a potential future failure. We performed a flurry of experiments to verify the accuracy of our model and demonstrate it outperforms state-of-the-art models in this task. Overall, the method is robust and intuitive, making it a useful tool for real-world applications.