Efficient and responsible fault detection in AUVs: SHAP-Interpretable cubic SVM on time-series data

Abstract

Fault detection in Autonomous Underwater Vehicles (AUVs) is crucial for ensuring their safe and efficient operation, especially in challenging underwater environments. In this study, we investigate traditional machine learning models to detect common AUV conditions, Add Weight, Pressure Gain Constant, PropellerDamageBad, and PropellerDamageSlight, using a comprehensive dataset of sensor readings and control signals. First, the data were segmented into four time windows (20, 50, 100, and 150 s) to address dimensionality and optimize model performance. To complement these time-series features, we then extracted basic descriptive statistical features (mean, standard deviation, min/max, 25th/50th/75th percentiles, skewness, and kurtosis) for each window and retrained our classifiers on this enriched feature set. Among all models, the Cubic SVM achieved the highest test accuracy of 96.7%. Finally, we applied SHapley Additive exPlanations (SHAP) to interpret the Cubic SVM’s predictions and identify the top-10 features driving correct classification for each class. Our results demonstrate that the Cubic SVM not only delivers a balanced trade-off between accuracy and computational efficiency, suitable for real-time, resource-constrained deployment, but also, through SHAP, yields actionable insights into the most discriminative sensor and control-signal statistics for each AUV condition. Future work will explore online adaptation and further model compression to enhance deployment in fully autonomous missions.