Data-Driven Propulsion System Fault Diagnosis for Deep-Sea Submersible

The propulsion system of the submersible is critical for deep-sea operations, yet traditional supervised fault diagnosis methods falter due to the scarcity and incompleteness of abnormal sensor data. To address this, we propose an unsupervised fault diagnosis method based on the GANomaly model, trained exclusively on normal operational data. Specifically, we introduce a temporal–spatial data transformation to preprocess multisensor time-series signals, converting them into 2-D representations that capture intersensor correlations and temporal dynamics. By learning healthy patterns from normal data only, the GANomaly approach enables anomaly detection through its reconstruction limitations on faulty data. Building on this capability, a dual-threshold technique using anomaly and component scores accurately identifies and localizes faults in both time and space. Tested on synthetic and real fault data from a five-year period of dives, this method achieves robust fault detection and localization with 93.6% accuracy, with results validated against on-site maintenance records.