Predicting a passenger ship's response during evasive maneuvers using Bayesian Learning
提出一个基于贝叶斯学习的元模型，利用大量六自由度数值模拟数据，预测客船在规避机动中的安全关键参数，精度达81%-98%，可辅助值班驾驶员决策。
• New meta-model predicting safety-critical parameters for a passenger ship when turning is presented;. • The meta-model considers a variety of operational and environmental parameters as well as wave stochastic realization;. • To train the meta-model on a large dataset, the Bayesian Learning algorithms were used;. • The dataset was developed using a high-fidelity 6DoF ship motion model ; . • The meta-model can be used for both forward and backward reasoning with high prediction power. The rapidly advancing automation of the maritime industry – for instance, through onboard Decision Support Systems (DSS) – can facilitate the introduction of advanced solutions supporting the process of collision avoidance at sea. Nevertheless, relevant solutions that aim to correctly predict a ship's behavior in irregular waves are only available to a limited extent by omitting the impact of wave stochastics on resulting evasive maneuvers. This is mainly due to the complexity of the phenomena, the existing couplings therein, and the time inefficacy in resolving the problem through real-time simulations. Therefore, this paper attempts to fill this knowledge gap by presenting a probabilistic, data-driven meta-model trained using an extensive set of 6DOF numerical simulations of vessel motions in irregular waves. For this purpose, machine learning adopting causal probabilistic modeling with Bayesian Belief Network (BBN) was employed. The latter offers two-way reasoning in the presence of uncertainty and provides insight into the meta-model's outcome. This, in turn, helps estimate a set of safety-critical parameters for a large passenger ship performing an evasive maneuver. This set comprises a huge quantity of ship turning circle parameters as well as the hull's rotational motions and resulting lateral accelerations, all simulated multiple times to consider the stochastic realization of the waves. The proposed meta-model can be used to assist watchkeeping officers’ decisions or raise their awareness concerning the possible consequences of evasive maneuvers performed. The achieved accuracy of the meta-model's prediction lies within a range from 81% to 98%, which makes it suitable for this purpose.