An identification modeling approach, locally weighted learning ( LWL) , was proposed for ship maneuve-ring motion modeling. First, samples were rearranged, and the input dimension was raised to solve the one-to-multi-ple mapping and inseparability of the ship motion states;second, a distance metric was trained for every sample by the leave-one-out cross validation;finally, the motion states of the ship were directly forecast by the weighted least squares in the local neighborhood. By constructing a local cost function, the defect that the conventional global opti-mal LWL algorithm was easily caught in the local optimality was avoided. Compared with the traditional mechanism modeling, the method settles the problem of parameter drift caused by multicollinearity and the unmodeled dynam-ics existing in the model. The algorithm realizes high-accuracy modeling for nonlinear systems by learning a group of artificial data and 3-DOF of the mariner class vessel. Compared with the back propagation neural network ( BPNN) prediction, the proposed scheme has improved generalization.%针对船舶操纵运动建模,本文提出了一种辨识建模方法,即局部最优的局部加权学习算法.该算法通过样本点重新排序和输入空间升维,解决了船舶运动状态一对多映射和不可分问题;并运用留一交叉验证为每个样本点训练一个距离测度,运用加权最小二乘在局部邻域中直接预测船舶操纵运动状态变量.构造局部目标函数,避免了传统的全局最优LWL算法容易陷入局部最优问题.与传统的机理建模相比,局部最优的局部加权学习算法克服了由多重共线性而引起的参数漂移和模型中存在未建模动态问题.通过一组人工数据和3自由度的Mariner轮的学习,实现了对非线性系统的高精度建模.与BPNN预报相比,具有较强的泛化能力.
展开▼
