基于CNN-LSTM的轨道车辆轮对运行状态识别方法研究  

Research on Recognition Method of Running State of Rail Vehicle Wheelset Based on CNN-LSTM

作　　者：雍心剑 陆正刚 Yong Xinjian;Lu Zhenggang(Insititute of Rail Transit,Tongji University,Shanghai 200092,China)

机构地区：[1]同济大学铁道与城市轨道交通研究院,上海200092

出　　处：《机电工程技术》2023年第5期19-24,76,共7页Mechanical & Electrical Engineering Technology

摘　　要：轨道车辆的轮对运行实时状态是评价车辆运行安全性及轮对导向控制的关键信息,针对轨道车辆在运行过程中轮对横移、摇头角实时状态直接测量成本高、难度大等问题,提出了一种基于卷积神经网络(CNN)内嵌长短记忆(LSTM)网络的轮对运动状态识别及预测方法。以轮对横向加速度、轮对摇头角加速度、一系悬挂位移量等易测信号构成特征集,通过CNN对时序信号进行多维度空间特征提取并输入到LSTM中捕获时序特征,最后通过全连接层输出轮对横移及摇头角的预测值,结合车辆运行的实际工况特点,对预测模型的泛化性及鲁棒性进行检验。仿真结果表明:相较于传统的单LSTM识别模型,CNN-LSTM模型能有效降低轮对运动状态的识别误差,且在不同运行工况以及车辆物理参数变化的情况下,该模型具有高鲁棒性,能够保持较高的预测精度。The real-time running state of wheelset of rail vehicles is the key information to evaluate running safety of vehicle and the wheelset guidance control.Aiming at the problems of high cost and difficulty in direct measurement of wheelset lateral movement and yaw angle parameters of rail vehicles during operation,a method of wheelset state recognition based on convolution neural network(CNN)embedded long and short term memory(LSTM)network was proposed.The feature set was composed of easy-to-measure signals such as the lateral acceleration of the rail vehicle wheel set,the angular acceleration of the wheel set shaking head,and the primary suspension displacement.The multi-dimensional spatial feature extraction of the time series signal was carried out through CNN and input to LSTM to capture the time series feature.Finally,the predicted values of the wheel set lateral displacement and shaking head angle were output through the full connection layer,and the generalization and robustness of the prediction model were tested in combination with the actual operating conditions of the vehicle.The results show that compared with the traditional single LSTM recognition model,the CNN-LSTM model can effectively reduce the recognition error of the wheel set motion state,and the model can still maintain high prediction accuracy under different conditions and when the vehicle physical parameters change.

关 键 词：轨道车辆 轮对运动状态 状态识别 卷积神经网络 长短时记忆网络 

分 类 号：TM911[电气工程—电力电子与电力传动]