Research on AUV navigation state prediction method using multihead attention mechanism in a CNN-BiLSTM model

In response to the issue of low accuracy in predicting the navigation state of existing AUV motion models, a predictive model for AUV navigation state based on a combination of a time-domain convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) recurrent neural network enhanced by a multi-head attention mechanism is proposed. Initially, the CNN extracts the feature sequence of the AUV navigation state. Subsequently, the BiLSTM network is utilized to learn the trajectory sequences of preceding and following conditions to extract dependencies within each navigation state in the sequence. The multi-head attention mechanism is then introduced to adjust the weights of different attribute features, prioritizing those with a greater impact on the navigation state. This ultimately facilitates the accurate prediction of the AUV navigation state. A comparative experiment using AUV lake test data demonstrated that the CNN-MABiLSTM model achieves higher prediction accuracy and better fitting than the CNN-BiLSTM model, thus verifying the effectiveness and practicality of the designed CNN-MAiLSTM model in predicting AUV navigation states.