Real-Time Fault Detection in Ocean Observing Platforms

Ocean observing platforms (OOPs) comprising moored buoys, gliders, autonomous surface vehicles, and fixed coastal stations - provide untold value for climate science, marine safety, and resource management. Persistence in the harsh marine environment exposes both sensors and platforms to mechanical degradation, biofouling, communication loss, and power faults, which decrease data quality and operational uptime. This article describes a full-fledged approach to realtime fault detection, developed specifically for OOPs, that combines edge-based pre-processing, multi-sensor fusion, statistically-based anomaly detection, and lightweight machinelearning classifiers designed for constrained computing environments. Robustness to missing data, asymmetric sensor reliabilities, and intermittent communications have been prioritized in the pipeline. The benchmarked approach has demonstrated faster detection latency, higher true positive rates (TPR), and lower rates of false alarms within realistic disturbance scenarios, versus traditional threshold-and-rule, or centralized post hoc analytics approaches. Guidelines are provided for implementation covering data conditioning, target selection (e.g., which models to choose given constrained compute footprints), adaptive thresholds, and remote update mechanisms.