Real-time prediction of full-scale ship maneuvering motions at sea under random rudder actions based on BiLSTM-SAT hybrid method

The prompt identification and prediction of ship maneuvering motions under random rudder actions are crucial for providing valuable navigation decisions in practical navigations. In this study, a hybrid modeling method (BiLSTM-SAT) combining bidirectional long short-term memory (Bi-LSTM) and scaled dot-product attention (SAT) mechanism is developed to adaptively capture the time-series dynamic features of the ship system with multiple degrees of freedom (DOF) and to predict the full-scale ship maneuvering motion at sea in real time. Firstly, the ability of the identified model by BiLSTM-SAT method to predict the 3-DOF nonstandard maneuvering motion of an unmanned surface vessel (USV) in model scale under random rudder actions is validated. On this basis, utilizing the ship motion data from sea trials, the developed BiLSTM-SAT method is applied to predict the time-series 5-DOF maneuvering motions for a full-scale YUKUN ship under the impacts of environmental disturbances and random rudder actions. The results demonstrate that comparing with the traditional LSTM and back propagation (BP) neural network methods, BiLSTM-SAT method can more accurately and stably predict the full-scale ship maneuvering motions in real time characterized by coupled nonlinearity and stochasticity features under variable environmental impacts and random rudder actions with satisfactory confidence level.