Deep Learning-Based Non-Parametric System Identification and Interpretability Analysis for Improving Ship Motion Prediction

The proposed hybrid model integrates a convolutional neural network, bidirectional long short-term memory network, and attention mechanism. This model is applied to the nonparametric system identification of ship motion, incorporating wind factors. The model processes input data with different historical dimensions after preprocessing, extracts local features using a CNN layer, captures bidirectional temporal dependencies via a BiLSTM layer to provide comprehensive bidirectional information, and finally introduces a multi-head attention mechanism to enhance the model’s expressive and learning capabilities. However, the use of deep neural networks introduces difficulties in explaining internal mechanisms. The coupled CNN-BiLSTM-Attention model with SHapley Additive exPlanations was adopted for the prediction of ship motion processes and the identification of key input feature factors. The effectiveness of the proposed model was validated through experiments using a ship free-running motion dataset with wind interference. The findings indicate that, in comparison to conventional single-architecture models and composite architecture models, the proposed model attains smaller prediction errors and demonstrates augmented generalizability and robustness.