论文部分内容阅读
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一化处理后,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,NARX—IR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。
The coupling dynamic model of bow and mesh was established. The dynamic simulation of this model was carried out by using MATLAB software Simulink to obtain the contact line surface irregularity and contact force data. The contact line surface irregularity and bow mesh contact force data were normalized After processing, they are respectively used as input and output of NARX neural network. The traditional Bayesian regularization algorithm is improved, and the modified Bayesian regularization algorithm is used to modify the weights of NARX neural network. Improved NARX (NARX-IR) neural network method; NARX-IR neural network method for contact line surface irregularities and bow contact force analysis. Based on the root mean square error and the correlation coefficient, the BP neural network based on the LM algorithm, the NARX-BR neural network based on the traditional Bayesian regularization algorithm and the NARX-IR neural network method Performance evaluation. The results show that the BP-LM neural network method is difficult to describe the complicated relationship between contact line surface irregularity and bow contact force. The root mean square error of NARX-IR neural network method is less than that of NARX-BR nerve in both training and prediction Network method, and the correlation coefficient is greater than the NARX-BR neural network method. It can be inferred that the NARX-IR neural network method is more suitable for the analysis of the relationship between the contact surface irregularity and bow contact force.