Spatio-Temporal Graph Neural Network for Autonomous Anomaly Detection in Underwater Glider Telemetry

Underwater gliders are vital for sustained ocean monitoring, yet limited satellite communication and reliance on skilled pilots constrain fleet scalability. Automated anomaly detection can reduce operator burden and improve mission reliability. To address this challenge, this study proposes a spatio-temporal graph neural network (STGNN) adapted for underwater glider telemetry. The model treats each sensor as a graph node and couples a fixed intersensor structure with temporal dynamics to detect anomalous windows within the sensor channels. Its performance is then contextualized through a comparative evaluation against deep and classical baselines under a pooled cross-mission protocol with configurable manual mission splits, together with principled threshold calibration through quantile, rolling-statistics, and extreme value strategies. It also has explicit latency and false-alarm reporting. Experiments conducted on Slocum G2 deployments with synthetic fault injection show that the proposed STGNN is the strongest overall model in the evaluation suite with an F1 score of 0.986. The codebase and evaluation protocol are provided for general use.