For collecting the test data of catenary irregularities and contact forces, the pantograph/catenary dynamics model was established and dynamic simulation was conducted using MATLAB/Simulink. After normalization, the catenary irregularities data and contact forces data were used respectively as the inputs and outputs of Nonlinear Auto-Regressive with eXogenous Input (NARX) neural networks. The traditional Bayesian Regularization (BR) was improved to modify the weight of NARX neural network, and the improved NARX (NARX-IR) neural network method was obtained. With NARX-IR neural network method, the correlation analysis between catenary irregularities and pantograph-catenary contact forces was carried through. The BP neural network based on LM algorithm (BP-LM), the NARX neural network based on traditional Bayesian regularization algorithm (NARX-BR), and the NARX neural network based on improving traditional Bayesian regularization algorithm (NARX-IR), three methods were tested and compared, and root mean square error and correlation coefficient were used for performance evaluation. The result shows that it's difficult for BP-LM neural network method to describe the complicated correlation between catenary irregularities and pantograph-catenary contact forces. Whether in training or predicting, the root mean square error of NARX-IR neural network method is lower than that of NARX-BR, and its correlation coefficient is greater than that of NARX-BR It is concluded that the proposed NARX-IR neural network method is more suitable for analyzing the correlation between catenary irregularities and pantograph-catenary contact forces.%建立弓网耦合动力学模型,采用软件MATLAB的Simulink模块对该模型进行动态仿真,获取接触线表面不平顺和弓网接触力数据；对接触线表面不平顺和弓网接触力数据进行归—化处理后,分别作为非线性自回归(NARX)神经网络的输入和输出；对传统的贝叶斯正则化算法进行改进,并采用改进的贝叶斯正则化算法进行NARX神经网络权值修正,得到改进的NARX(NARX-IR)神经网络方法；利用NARX-IR神经网络方法进行接触线表面不平顺与弓网接触力的关联分析.采用根均方误差和相关系数,对基于LM算法的BP(BP-LM)神经网络方法、基于传统贝叶斯正则化算法的NARX(NARX-BR)神经网络方法和NARX-IR神经网络方法进行性能评价.结果表明:BP-LM神经网络方法难以描述接触线表面不平顺与弓网接触力的复杂关联关系；不论在训练还是预测中,NARXIR神经网络方法的根均方误差均小于NARX-BR神经网络方法,而相关系数则大于NARX-BR神经网络方法.由此可推断:NARX-IR神经网络方法更适合于分析接触线表面不平顺与弓网接触力的关联关系.
展开▼
