Anomaly Detection and Fault Diagnostics for Underwater Gliders Using Deep Learning

Abstract

Near Real-Time (NRT) anomaly detection and fault diagnostics for underwater gliders are challenging because satellite connections with limited bandwidth allow only decimated data to be sent back from the remote vehicle, whilst on-board systems are constrained by power and computational limits. Currently, anomaly detection and fault diagnostics for such vehicles require intensive monitoring from the operating pilots, which prohibits large scale deployments with multiple vehicles. This paper presents a system with NRT anomaly detection and fault diagnostics for multi-vehicle underwater glider fleets based on Bidirectional Generative Adversarial Networks with assistive hints. The unsupervised anomaly detection system is applied to assist in annotating deployment datasets to train a supervised fault diagnostics model. The results suggest that the anomaly detection system has successfully detected different types of anomalies, validated against model-based and rule-based approaches. In addition, the supervised fault diagnostics system has achieved high fault diagnostics accuracy on the test dataset.