Interpretable machine learning approach for AUV motion prediction based on lake-test data

Interpretable models are essential for deploying deep-learning techniques in marine activities. However, the layered complexity of state-of-the-art deep-learning architectures hinders mechanistic insight and limits adoption. Here, we introduce a three-degree-of-freedom surrogate model that renders a multivariate long short-term memory network with multiple input and output channels (MIMO-LSTM) for autonomous underwater vehicles (AUVs) transparent and tractable. The surrogate is built on a least-squares support-vector machine selected for its superior approximation and generalization capacity. Benchmark manoeuvres show that the surrogate retains predictive fidelity up to 95.8% while accelerating inference by 98.6% relative to the parent MIMO-LSTM. Interpretable parameter-dependence plots quantify the contribution of individual variables and deliver mechanistically grounded, transparent forecasts of AUV manoeuvring dynamics. • A 3-DOF surrogate model based on LS-SVM is proposed to enhance the interpretability of MIMO-LSTM for AUVs. • The surrogate model increases prediction speed by 98.6% while maintaining acceptable accuracy, outperforming the original MIMO-LSTM. • Interpretable parameter-dependence plots quantify variable contributions to AUV maneuvering dynamics predictions, ensuring transparent results. • The proposed framework is verified by experiments based on an AUV model.