A remote anomaly detection system for Slocum underwater gliders

Abstract Marine Autonomous Systems (MAS) operating at sea beyond visual line of sight need to be self-reliant, as any malfunction could lead to loss or pose a risk to other sea users. In the absence of fully automated on-board control and fault detection tools, MAS are piloted and monitored by experts, resulting in high operational costs and limiting the scale of observational fleets that can be deployed simultaneously. Hence, an effective anomaly detection system is fundamental to increase fleet capacity and reliability. In this study, an on-line, remote fault detection system is developed for underwater gliders. Two alternative methods are analysed using time series data: feedforward deep neural networks estimating the glider’s vertical velocity and an autoencoder. The systems are trained using field data from four baseline deployments of Slocum gliders and tested on six deployments of vehicles suffering from adverse behaviour. The methods are able to successfully detect a range of anomalies in the near real time data streams, whilst being able to generalise to different glider configurations. The autoencoder’s error in reconstructing the original signals is the clearest indicator of anomalies. Thus, the autoencoder is a prime candidate to be included into an all-encompassing condition monitoring system for MAS.