Improved Prediction Model of Metro Wheel Wear Based on BiLSTM Network
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
万方数据
针对地铁车轮磨耗数据时间跨度较长引起的长期依赖问题,为了进一步提升预测精度,提出一种将麻雀搜索算法(sparrow search algorithm,SSA)优化双向长短期记忆网络(bidirectional long short term memory,BiLSTM)的改进BiLSTM(SSA-BiLSTM)网络模型,用于地铁车轮磨耗预测.首先,利用麻雀搜索算法对双向长短期记忆网络算法的神经元个数、迭代次数、输入批量和学习率等超参数在给定范围内进行寻优,得到参数最优值;然后,以参数最优值来构建改进BiLSTM网络模型,对车轮磨耗进行预测分析;最后,以车轮踏面磨耗和轮缘磨耗作为研究对象,将某地铁1车厢1号车轮的现场实测历史磨耗数据作为输入,对该模型进行训练及验证分析,并与多层感知机(multilayer perceptron,MLP)、LSTM、BiLSTM以及SSA-LSTM模型的预测结果进行对比.研究结果表明:SSA-Bi-LSTM模型的车轮磨耗预测精度更高,与LSTM、BiLSTM以及SSA-LSTM网络模型相比,踏面磨耗的平均绝对百分误差(mean absolute percentage error,MAPE)分别降低了 13.28％、10.32％、1.47％,轮缘磨耗分别降低了 9.5％、0.46％、0.02％;分别对同一地铁2号、4号车厢的1号位置车轮磨耗进行预测,并与磨耗实测数据进行对比,踏面磨耗的平均绝对百分比误差分别为1.34％、1.42％,轮缘磨耗的平均绝对百分比误差分别为0.18％、0.19％,验证了本文所提模型具有良好的泛化性,为地铁轮对智能化管理提供理论支持,延长车轮使用寿命.
In order to address the issue of long-term dependence caused by the extended time span of wheel wear data and improve the prediction accuracy,an improved BiLSTM metro wheel wear prediction model is proposed by optimizing Bidirectional long short-term memory network(Bi-LSTM)with Sparrow search algorithm(SSA).Firstly,the hyperparameters of the Bi-LSTM algorithm,such as the number of neurons,iteration count,input batch size,and learning rate,are optimized using the SSA.This optimization process is conducted within a specified range to obtain the optimal values of these hyperparameters.This optimization process aims to obtain the optimal parameter values.Subsequently,the SSA-BiLSTM network model is constructed using these optimal parameter values to predict and analyze wheel wear.Tread wear and flange wear are taken as the research objects,and the measured historical wear data of wheel No.1 of the metro's carriage # 1 are used as inputs to metro and validate the model,and compare the prediction results with those of MLP,LSTM,BiLSTM and SSA-LSTM models.The results show that the improved bidirectional long short-term memory network model has higher wear prediction accuracy,and the mean absolute percentage error(MAPE)of tread wear is reduced by 13.28％,10.32％,and 1.47％,and flange wear by 9.5％,0.46％,and 0.02％.The wear of the No.1 wheel of the same metro No.2 and No.4 cars is predicted and compared with the measured wear data.The average absolute percentage error of tread wear is 1.34％and 1.42％,respectively,and the average absolute percentage error of rim wear is 0.18％and 0.19％,respectively.The results confirm that the model exhibits strong generalization capabilities.The wheel wear prediction model based on improved BiLSTM network(SSA-Bi-LSTM)has high prediction accuracy and good generalization,which provides theoretical support for the intelligent management of metro wheelsets and prolongs wheel service life.
metrowear predictionsparrow search algorithmbidirectional long short time memory networks
NETL
NSTL
万方数据
