基于未来可预知特征的高速列车变流器超温故障KFF-LSTM预测方法  

KFF-LSTM high temperature fault prediction method using in high-speed rail train converter based on known future features

作　　者：刘典 秦勇[1] 杨伟君[2] 周伟[4,5] 扈海军[2] 杨宁[2] 刘冰[2] 赵鹏飞 董光磊 LIU Dian;QIN Yong;YANG Weijun;ZHOU Wei;HU Haijun;YANG Ning;LIU Bing;ZHAO Pengfei;DONG Guanglei(The State Key Laboratory of Rail Traffic Control and Safety,Beijing Jiaotong University,Beijing 100044,China;Locomotive&Car Research Institute,China Academy of Railway Sciences Group Co.Ltd.,Beijing 100081,China;Beijing Zongheng Electro-Mechanical Technology Co.Ltd.,Beijing 100094;School of Traffic&Transportation Engineering,Central South University,Changsha 410075,China;Key Laboratory of Traffic Safety on Track of Ministry of Education,Changsha 410075,China)

机构地区：[1]北京交通大学轨道交通控制与安全国家重点实验室,北京100044 [2]中国铁道科学研究院机车车辆研究所,北京100081 [3]北京纵横机电科技有限公司,北京100094 [4]中南大学交通运输工程学院,湖南长沙410075 [5]中南大学轨道交通安全教育部重点实验室,湖南长沙410075

出　　处：《中南大学学报（自然科学版）》2023年第8期3370-3378,共9页Journal of Central South University:Science and Technology

基　　金：国家重点研发计划项目(2019YFB1705000)。

摘　　要：提出一种基于未来可预知特征的长短期记忆网络(KFF-LSTM)长时序数据故障预测模型,引入包括动态物理量、缓变/静态物理量与恒定状态量等未来可预知特征数据进行训练,同时通过调整经典LSTM模型架构的输出门、状态更新单元以及隐变量以优化模型的预测准确率与及时性;最后,基于高速动车组线路运行的牵引变流器电机定子温度数据,分别以平均绝对误差(MAE)、均方根误差(RMSE)和R2分数表征回归预测准确度,以偏滞步数表征响应及时性的评价指标,对KFF-LSTM、RNN、GRU及LSTM预测模型进行消融实验及可视化对比。研究结果表明,相对于其他3种方法,提出的KFF-LSTM方法在测试集超前1~16步预测的MAE和RMSE最高可分别降低18.0%和10.8%,R2分数提升可达26.5%,在超前16步的预测偏滞步数要优于其他方法40.0%,在高速列车的长序列数据故障预警场景中具有较好的应用推广前景。A long sequence fault prediction model was proposed based on known future feature LSTM method(KFF-LSTM),which utilizes predictable future features including dynamic,gradual/static and constant state for model training.At the same time,the output gate,state updating unit and hidden variable in the classical LSTM model were coordinated to optimize its prediction accuracy and time delay.Finally,the model was verified by the stator temperature data of traction converter motor on high-speed EMU site.Ablation test was conducted and visually compared by KFF-LSTM,RNN,GRU and LSTM models in terms of the mean absolute error(MAE),root mean square error(RMSE),and R2-score to demonstrate the evaluation accuracy,and delay steps to represent timeliness.The results show that compared with the other three models,the MAE and RMSE error reduction of the proposed KFF-LSTM best prediction can reach 18.0%and 10.8%,respectively,the R2-score increasing can reach 26.5%in 1-step to 16-step ahead.The time delay is also optimized by 40.0%in 16-step prediction.It has a good application and promotion prospect in the high-speed railway long-sequence data fault early warning scene.

关 键 词：LSTM 未来可预知特征 长时间序列 数据驱动 故障预警 

分 类 号：U279.3[机械工程—车辆工程]