基于GWO-CNN-LSTM的铁路轨道高低不平顺值反演模型研究

Study on Inversion Model of Railway Track Longitudinal Irregularity Value Based on GWO-CNN-LSTM

为利用列车的车体振动加速度来准确反演铁路的轨道高低不平顺值,提出一种基于灰狼优化算法(GWO)、卷积神经网络(CNN)和长短期记忆网络(LSTM)构建车体振动加速度与轨道高低不平顺值的关系模型(GWO-CNN-LSTM)。首先,将轨检车的实测数据根据轨道检测数据特点,采用莱因达准则进行异常值剔除的预处理;然后,利用处理后的数据以车体振动加速度作为模型的输入,以轨道高低不平顺值作为模型的输出,利用CNN学习车体振动加速度的波形信息,将CNN学习到的特征输入LSTM;最后,再用GWO优化LSTM模型的关键参数,进而反演出轨道高低不平顺值,为突出该模型的适应性,又随机选取了其他区段进行了反演。文中采用3个性能指标来评价模型,并与其他经典的方法进行对比分析。结果表明,GWO-CNN-LSTM模型的均方根误差、平均绝对误差和拟合优度分别为0.141、0.107和0.977,与单一的LSTM相比,GWO-CNN结合LSTM可以使拟合优度提高20.1%;与循环神经网络、BP神经网络和支持向量回归相比,所提出的GWO-CNN-LSTM模型其均方根误差降低68.1%~79.1%,平均绝对误差降低60.4%~68.3%,拟合优度提高27.7%~44.9%,验证了GWO-CNN-LSTM模型用于轨道高低不平顺值反演的有效性。GWO-CNN-LSTM模型为铁路轨道高低不平顺反演研究提供了新的思路。

In order to use the acceleration of the train body vibration to accurately inverse the track longitudinal irregularity value of railroads,this paper proposes a model(GWO-CNN-LSTM)based on Gray Wolf Optimization algorithm(GWO),Convolutional Neural Network(CNN)and Long Short Term Memory Network(LSTM)to construct the relationship between vehicle body vibration acceleration and track longitudinal irregularity value.Firstly,the measured data of the track inspection vehicle are pre-processed according to the characteristics of the data using the PauTa criterion for outlier rejection,and then the processed data are used with the vehicle vibration acceleration as the input of the model in which the track longitudinal irregularity is the output.The CNN is used to learn the waveform information of the vehicle body vibration acceleration,and the features learned by the CNN are input to the LSTM.Finally the key parameters of the LSTM model is optimized with GWO,then the inverse perform of the track longitudinal irregularity value is made.To highlight the adaptability of the model,additional zones are randomly selected for inversion.Three performance metrics are used to evaluate the model and compare with other classical methods.The results show that the root mean square error,the mean absolute error and the goodness of fit of the GWO-CNN-LSTM model are 0.141,0.107 and 0.977 respectively,and the GWO-CNN combined with LSTM can improve the goodness of fit by 20.1%compared with a single LSTM;compared with recurrent neural network,BP neural network and support vector regression,the proposed GWO-CNN-LSTM model has 68.1%~79.1%lower root mean square error,60.4%~68.3%lower mean absolute error,and 27.7%~44.9%higher goodness of fit,which verifies the GWO-CNN-LSTM model for inversion of track longitudinal irregularity values validity.The model provides a new idea for the study of railroad track longitudinal irregularity value inversion.