Physics-Informed Ensemble Learning for Detection of Malfunctioning Ocean Bottom Pressure Gauges: A Case Study on the S-Net Network

In this study, we analyse bottom pressure time series from $139$ operational stations of the S-net network to detect potential anomalies. As of March 12, 2025, $15$ stations were identified as malfunctioning: S1N09, S1N13, S1N21, S2N06, S2N18, S2N20, S2N25, S3N02, S3N04, S3N22, S4N10, S4N24, S5N10, S5N12, S5N16. Notably, the failure of station S2N25 was captured in real time during observations. The analysis revealed that increased depth is statistically associated with a higher risk of sensor malfunction (p < 0.05). To enable automatic identification of malfunctioning pressure gauges, we introduce an ensemble learning model that combines five base predictors with a random forest meta-classifier. The model aggregates prior knowledge of the physical characteristics of the signal with machine learning to classify stations as either operational or malfunctioning. The central assumption is that functional sensors should display a clear tidal signal, whereas malfunctioning sensors produce irregular, noise-like outputs. The average F1-score on the test set reached 0.91 for the malfunctioning class and 0.99 for the functional class. The full project code is provided on GitHub: https://github.com/Olegg66/SNET-malfunc