Stern pressure identification using CNN-BiLSTM model based on bayesian optimisation and attention mechanism

Accurate identification of the stern pressure field holds significant engineering implications for controlling hull structure vibrations and propeller excitation effects. In order to address the accuracy defects of the traditional distributed load inversion method when applied to complex surface structures, a baseline neural network model has been introduced and improved. Furthermore, a method of intelligent identification of stern pressure based on the Bayesian optimization algorithm and the attention mechanism of the CNN- BiLSTM model has been proposed. Taking the KCS container ship as the research object, we constructed a multi-physical-field coupled dataset containing 65 stern pressure measurement points and 9 structural strain measurement points through CFD simulations under five propeller rotational speed conditions. CNN and BiLSTM are used to extract spatio-temporal features from the dataset. Dynamic feature weight allocation is achieved through an attention mechanism, while Bayesian optimisation determines hyperparameter values to reduce bias. Results demonstrate that compared with basic models (CNN-BiLSTM, GRU, etc.), BOA-CNN-BiLSTM achieves optimal performance across MAE, MSE, RMSE, and MAPE metrics, with R 2 reaching 0.987. This method achieves high-precision and high-reliability reconstruction of multi-point pressure data at the stern through finite strain monitoring, providing an effective solution for distributed load identification in complex curved structures.