SeaBil: Self-attention-weighted ultrashort-term deep learning prediction of ship maneuvering motion

Abstract

Accurate prediction of motion dynamics fundamentally promotes the autonomy of intelligent ships, but faces great challenges in modeling mechanism. In this paper, to establish data-driven recurrent mapping within ship motion dynamics, an ultrashort-term deep learning predictor is innovatively developed by elaborately creating a self-attention-weighted bidirectional long short-term memory (Bi-LSTM) network in conjunction with 1-dimensional convolution (Conv-1D), named SeaBil. To be specific, combined with the sliding-window technique, the Conv-1D is devised to convert 5-dimensional-input samples consisting of course angle, yaw rate, roll angle, total speed and rudder angle into 1-D-feature vectors, thereby extracting coupled feature maps within the current data-window. The Bi-LSTM is further deployed to recurrently learn forward and reverse feature maps of ship motion time-series data. Self-attention mechanism cascaded in the serial is employed to adaptively assign time-varying weights of sample instants within the current window such that the ultrashort-term prediction of course angle, yaw rate, roll angle and total speed can be achieved. Eventually, comprehensive comparisons to typical approaches using real-world samples demonstrate the superiority of SeaBil in terms of ultrashort-term prediction accuracy.