摘要
轨道车辆实际线路运行中振动状态是评价车辆运行稳定性、乘坐舒适性及安全运营的重要信息,但运行时面临多变工况条件会造成轨道车辆振动状态实时测量困难的问题。针对这一挑战,提出一种融合贝叶斯优化算法(BO)、卷积神经网络(CNN)和长短期记忆网络(LSTM)的车辆振动状态识别新方法。通过对测量得到的振动相关数据进行归一化预处理;利用处理后的数据以转向架横向速度及横摆角速度作为模型的输入,以车体横向加速度及横摆角加速度作为模型的输出,采用CNN学习并提取转向架横向速度及横摆角速度的波形特征,将相应的特征输入LSTM;同时利用BO优化得到LSTM模型的最优超参数配置,进而输出车体横向加速度及横摆角加速度,实现实时状态识别轨道车辆车体横向振动状态。考虑到轨道车辆实际运行的工况特点,采用3个评价指标来分析模型有效性,并与其他经典的方法进行对比分析。研究结果表明,相较于传统的BP神经网络和CNN-LSTM模型,通过BO-CNN-LSTM模型实时识别车体横向振动状态能有效降低误差,准确性最高,并且功率谱密度波形也最为接近。面对不同车速及载荷等工况条件,模型的平均绝对误差和均方根误差分别在0.019和0.029以下,拟合度均在0.99以上,验证了BO-CNN-LSTM模型用于轨道车辆车体横向振动状态识别的有效性。研究结果为轨道车辆运行状态实时识别的研究提供了新思路。
The vibration state of rail vehicles in actual line operation is important information for evaluating the stability,ride comfort and safe operation of vehicles.However,it is difficult to measure the real-time vibration state of rail vehicles due to the variable operating conditions.To solve this challenge,a new method of vehicle vibration state identification based on Bayesian optimization algorithm(BO),convolutional neural network(CNN)and long short-term memory network(LSTM)is proposed.The measured vibration-related data are preprocessed by normalization.The lateral velocity and yaw velocity of the bogie in the processed data are used as inputs to the model,and the lateral acceleration and yaw acceleration of the vehicle body are taken as outputs of the model.The CNN is utilized to learn and extract the waveform characteristics of lateral velocity and yaw velocity of the bogie,and the corresponding characteristics are input into LSTM.Meanwhile,the BO optimization is adopted to obtain the optimal hyperparameter configuration of the LSTM model.Then,the lateral acceleration and yaw acceleration of the vehicle body are output to achieve the real-time state identification of the lateral vibration state of the rail vehicle body.Considering the characteristics of the actual operating conditions of rail vehicles,three performance indicators are applied to analyze the effectiveness of the model.Moreover,it is compared with other classical methods.The results show that compared with the traditional BP neural network and CNN-LSTM model,the real-time identification of the lateral vibration state of the vehicle body through the BO-CNN-LSTM model can effectively reduce the error,attain the highest accuracy and exhibit spectral density waveforms closest to the true values.In the face of operating conditions such as different vehicle speeds and loads,the average absolute error and root mean square error of the model are below 0.019 and 0.029 respectively.Additionally,the coefficient of determination is above 0.99.This verifies the effectiveness of the BO-CNN-LSTM model for the lateral vibration state identification of the vehicle body.The research results provide new perspectives for the real-time identification of the operational state of rail vehicles.
作者
周军超
刘乙翰
陈奥
章杰
ZHOU Junchao;LIU Yihan;CHEN Ao;ZHANG Jie(School of Mechanical Engineering,Sichuan University of Science&Engineering,Zigong 643000,China;College of Automotive and Mechanical Engineering,Changsha University of Science&Technology,Changsha 410114,China;Intelligent Policing Key Laboratory of Sichuan Province,Sichuan Police College,Luzhou 646000,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2024年第9期3900-3910,共11页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(12002123)
智能警务四川省重点实验室开放课题资助项目(ZNJW2023KFQN005)
泸州市科技计划资助项目(2023JYJ066)。
关键词
轨道车辆
车体横向振动
贝叶斯优化算法
卷积神经网络
长短期记忆网络
状态识别
rail vehicles
lateral vibration of vehicle body
Bayesian optimization algorithm
convolutional neural network
long short-term memory network
state identification