A novel interpretable intelligent attitude perception system for underwater vehicles based on pulsating pressure sensor array

Accurate perception of the sideslip angle of underwater vehicles relative to ocean currents is essential for precise motion control and effective early warning of complex hydrodynamic disturbances in marine environments. This study proposes, for the first time, an interpretable intelligent prediction method for sideslip angle based on a pulsating pressure sensor array. Wavelet decomposition is employed to extract multi-scale features from the pressure signals collected by the sensors, which are subsequently fed into a hybrid XGBoost–SHAP framework for the inversion and identification of the sideslip angle of an underwater vehicle. During model training, five-fold cross-validation combined with Bayesian optimization is adopted to enhance the generalization capability and improve hyperparameter tuning efficiency. Experimental results demonstrate that the proposed method achieves an identification accuracy of 91.77% for sideslip angle prediction. Moreover, under previously unseen flow velocity conditions, the model still maintains robust performance with an accuracy of 87.93%. In addition, the proposed framework quantifies and reveals the contribution magnitude and direction of wavelet-based multi-scale energy features to the model output, thereby providing interpretable insights into the model prediction mechanism. These findings provide a reliable solution for underwater flow sensing, offering valuable support for the intelligent perception and autonomous control of underwater vehicles.