基于ROC曲线优化的车辆行驶状态估计BP模型

BP neural network for vehicle state estimation was based on ROC curve optimization

针对现有行车状态估计器难以适应复杂非线性模型,结合BP神经网络在解决非线性系统方面表现出优良的性能,采用ROC曲线(受试者特征工作曲线)对BP神经网络算法进行优化,依据各个节点权重值的变化情况绘制学习机器相应的ROC曲线,将ROC曲线下方面积作为各个节点权重值选取的唯一准则,每次在同一节点进行变步长的搜索(大步长和小步长),并根据不同步长的搜索结果确定下一次步长的大小,以确定最佳的权重值,最后以波动性较强的车辆横摆角速度作为样例对算法进行验证.研究结果表明:通过ROC对其性能的评价,加速了BP网络的收敛速度,在一定程度上避免了出现局部最小值的情况,提高了模型的容错能力;优化后的模型在5%误判率的情况下有较高的击中概率,表现出更强的泛化能力,适应性更强.

In view of the inadequate adaptability of the existing traffic estimator to the complex nonlinear models, the receiver operating characteristic （ROC） curves were adopted to evaluate the algorithm of BP neural network which demonstrates excellent performance in solving problems with nonlinear systems. According to each node weight value, the corresponding ROC curve of the learning machine is drawn. The area of ROC curve is the only criterion for the selection of each node weights. The method is used for the searching of the minimum value through variable steps, i.e. maximum and small, at the same nodes. According to the results, the size of the next step for determining the best weights can be fixed. Finally, the algorithm verification can be carried out on the basis of the strong volatility data of vehicle yaw-rate. The results show that the evaluation of the performance of ROC may accelerate the convergence speed of the BP network to a certain extent which making it possible to avoid the local minimum. Furthermore, the fault tolerance model can be improved and the optimal model error probability of the hit rate goes higher in 5% of the cases. The optimized model showed a strong generalization capacity as well as the fine adaptability.