基于全局最优局部加权学习算法的船舶操纵运动辨识

Global-Optimal-Based Locally Weighted Learning for Ship Maneuvering Motion Identification

运用一种基于全局最优的局部加权学习(Locally Weighted Learning,LWL)算法进行船舶操纵运动辨识建模。该方法是一种基于计算机存储的离线学习的黑箱建模方法,直接考虑船舶运动状态输入与输出之间的映射关系,可克服传统机理建模及参数辨识模型中存在的参数漂移问题和未建模动态问题。对样本点进行重新排列并提高输入空间的维度,解决船舶运动状态一对多映射和不可分问题。通过学习"Mariner"轮的三自由度数学模型,并进行旋回试验、Z形试验及逆螺旋试验,验证算法的有效性。

Global-optimal-based Locally Weighted Learning （LWL） algorithm is applied to ship maneuvering motion identification modeling. LWL as a black box off-line modeling algorithm based on the computer memory, makes directly mapping between input and output of the ship motion states, therefore, eliminates the problems caused by parameter drifting and unmodeled dynamics, which exist in the mechanism modeling and parameter identification modeling. One-to-many mapping and inseparability of ship motion states are dealt with by sample rearrangement and raising the input dimension. The effectiveness of the algorithm is illustrated with learning a 3-D Mariner class vessel mathematical model then performing several maneuvering simulations, including： turning tests, zig-zag tests and reverse spiral tests.