基于多车型CNN-GRU性能预测模型的轨道状态评价

杨飞; 郝晓莉; 杨建; 孙宪夫; 高彦嵩; 张煜

doi:10.3969/j.issn.0258-2724.20211030

基于多车型CNN-GRU性能预测模型的轨道状态评价

doi: 10.3969/j.issn.0258-2724.20211030

杨飞^1,,
郝晓莉^2, ,,
杨建²,
孙宪夫^1,,
高彦嵩³,
张煜¹

1.
中国铁道科学研究院集团有限公司基础设施检测研究所，北京 100081
2.
北京交通大学电子信息工程学院，北京 100044
3.
中国国家铁路集团有限公司科技和信息化部，北京 100844

基金项目: 国家自然科学基金（61771042）；中国国家铁路集团有限公司科技研究开发计划（P2021T013）

详细信息

作者简介:
杨飞（1985—），男，副研究员，硕士，研究方向为轨道管理，E-mail：13811807268@163.com

通讯作者:
郝晓莉（1970—），女，副教授，博士，研究方向为信号处理，E-mail：xlhao@bjtu.edu.cn

中图分类号: U216.3
计量
- 文章访问数: 474
- HTML全文浏览量: 187
- PDF下载量: 76
- 被引次数: 0
出版历程
- 收稿日期: 2021-12-14
- 修回日期: 2022-06-16
- 网络出版日期: 2022-11-19
- 刊出日期: 2022-07-14

Track Condition Evaluation for Multi-vehicle Performance Prediction Model Based on Convolutional Neural Network and Gated Recurrent Unit

1.
Infrastructure Inspection Research Institute, China Academy of Railway Sciences Corporation Limited, Beijing 100081, China
2.
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
3.
Department of Science, Technology and Information Technology, China State Railway Group Co., Ltd., Beijing 100884, China

摘要

摘要:
不同车型高速综合检测列车的动力学传递特性不同，使得其对同一线路的车体加速度评价结果存在一定差异. 为解决上述问题，本文基于多列动检车的检测数据，将卷积神经网络（convolutional neural network，CNN）与门控循环单元（gated recurrent unit，GRU)相结合，建立了多车型车辆动力学响应预测模型，通过输入多项实测轨道不平顺和车速预测各车型的车体垂向和横向加速度，并将多车型车体加速度预测值的最大包络作为轨道状态评价依据. 结果表明：将高低、轨向不平顺等8项轨道不平顺和车速共同作为输入参数的模型预测性能最优，车体垂向和横向加速度预测的评估指标分别提升了5%～13%和25%～36%；CNN-GRU模型所预测的车体加速度在时域和频域均与实测结果吻合较好，相关系数最大达到0.902；且相比于BP （back propagation）神经网络，各项车体垂向和横向加速度预测的评估指标分别提升了36%～109%和11%～167%；针对某轨道几何状态不良区段应用效果，预测6种车型中有4种车型达到车体垂向加速度Ⅰ级或Ⅱ级超限，有1种车型达到车体横向加速度Ⅰ级超限，提高了轨道状态评价的准确性和一致性.
- 轨道不平顺 /
- 车体加速度 /
- 轨道状态评价 /
- 门控循环单元（GRU） /
- 卷积神经网络（CNN）
Abstract:
The dynamic transmission characteristics of different types of high-speed track inspection vehicles are different, which makes the evaluation results of vehicle body acceleration on the same railway line different. To solve the above problem, the convolutional neural network (CNN) is combined with the gated recurrent unit (GRU) to establish a dynamic response prediction model for multi-vehicle dynamic response, which predicts the vertical and lateral acceleration of each vehicle by inputting a number of measured track irregularities and vehicle speeds, and uses the maximum envelope of the predicted values of multi-vehicle acceleration as the basis for track state evaluation. The results show that the model with eight track irregularities and vehicle speed, such as longitudinal irregularity, horizontal irregularity, as input parameters has the best prediction performance, and the evaluation indices of vertical and lateral vehicle acceleration prediction are increased by 5%–13% and 25%–36%, respectively. The vehicle acceleration predicted by the CNN-GRU model is in good agreement with the measured results in both time domain and frequency domains, with the maximum correlation coefficient of 0.902. Compared with back propagation (BP) neural network, CNN-GRU improves the evaluation indices of vertical and lateral vehicle acceleration prediction by 36%–109% and 11%–167%, respectively. The application result in a section with poor track geometry state shows that four out of the six vehicle types reach the level Ⅰ or Ⅱ overrun of the vehicle vertical acceleration, and one vehicle type reaches the level Ⅰ overrun of the vehicle lateral acceleration, which improves the accuracy and consistency of the track state evaluation.
- track irregularity /
- vehicle body acceleration /
- track status evaluation /
- gated recurrent unit (GRU) /
- convolutional neural network (CNN)

HTML全文

图 1 一维CNN结构

Figure 1. Structure of one-dimensional CNN

下载: 全尺寸图片幻灯片

图 2 GRU单元结构

Figure 2. Unit structure of GRU

下载: 全尺寸图片幻灯片

图 3 CNN-GRU网络结构

Figure 3. Network structure of CNN-GRU

下载: 全尺寸图片幻灯片

图 4 超参数对车体加速度预测性能的影响

Figure 4. Influence of super parameters on prediction performance of vehicle body acceleration

下载: 全尺寸图片幻灯片

图 5 车速200 km/h的CRH2A-2010车预测与实测结果对比

Figure 5. Comparison between predicted and measured results of CRH2A-2010 at the speed of 200 km/h

下载: 全尺寸图片幻灯片

图 6 实测轨道不平顺

Figure 6. Measured track irregularity

下载: 全尺寸图片幻灯片

图 7 不同车型的CNN-GRU模型测试结果

Figure 7. Test results of CNN-GRU model for various track inspection vehicles

下载: 全尺寸图片幻灯片

表 1 轨道不平顺输入组合的CNN-GRU评价指标

Table 1. CNN-GRU evaluation index of track irregularity input combination

组号	输入	加速度	E_MAE/ （m·s⁻²）	E_RMSE/ （m·s⁻²）	U_TIC	ρ
①	车速 + 长波高低 + 长波轨向	垂向	0.078	0.098	0.291	0.839
①	车速 + 长波高低 + 长波轨向	横向	0.053	0.067	0.404	0.701
②	① + 高低 + 轨向	垂向	0.074	0.093	0.274	0.858
②	① + 高低 + 轨向	横向	0.049	0.062	0.389	0.737
③	② + 水平	垂向	0.069	0.087	0.259	0.875
③	② + 水平	横向	0.040	0.051	0.294	0.855
④	③ + 三角坑	垂向	0.070	0.088	0.257	0.875
④	③ + 三角坑	横向	0.041	0.053	0.312	0.831
⑤	④ + 超高	垂向	0.070	0.088	0.262	0.872
⑤	④ + 超高	横向	0.038	0.049	0.279	0.847
⑥	⑤ + 轨距	垂向	0.068	0.086	0.253	0.880
⑥	⑤ + 轨距	横向	0.034	0.044	0.256	0.879
⑦	⑥ （去除车速）	垂向	0.071	0.089	0.264	0.869
⑦	⑥ （去除车速）	横向	0.038	0.049	0.289	0.848

下载: 导出CSV

表 2 不同车型的训练集和测试集里程

Table 2. Kilometrages of training set and test set for various track inspection vehicles km

序号	车型	训练集总里程	测试集总里程
1	CRH2A-2010	400 （上行）	200 （上行）
2	CRH2C-2150	840 （上行）	200 （下行）
3	CRH380BJ-A-0504	400 （上行）	210 （上行）
4	CRH5J-0501	400 （下行）	100 （下行）
5	CRH380AJ-0201	940 （上行）	200 （上行）
6	CRH380BJ-0301	200 （下行）	100 （上行）

下载: 导出CSV

表 3 小波分解层及波长范围

Table 3. Wavelet decomposition layer and wavelength range

小波层	D1	D2	D3	D4	D5	D6	D7	D8	A8
波长范围/m	(0.5, 1.0]	(1.0, 2.0]	(2.0, 4.0]	(4.0, 8.0]	(8.0, 16.0]	(16.0, 32.0]	(32.0, 64.0]	(64.0, 128.0]	>128.0

下载: 导出CSV

表 4 多车型的不同模型评估指标对比

Table 4. Comparison of the evaluation index of different models for multi-vehicle

模型	车型	加速度	E_MAE/ (m·s⁻²)	E_RMSE/ (m·s⁻²)	U_TIC	ρ
BP	CRH2A-2010	垂向	0.113	0.152	0.692	0.214
	CRH2A-2010	横向	0.059	0.078	0.698	0.233
	CRH2C-2150	垂向	0.118	0.151	0.716	0.343
	CRH2C-2150	横向	0.054	0.071	0.693	0.345
	CRH380BJ-A-0504	垂向	0.064	0.081	0.618	0.468
	CRH380BJ-A-0504	横向	0.060	0.064	0.752	0.161
	CRH5J-0501	垂向	0.051	0.066	0.633	0.440
	CRH5J-0501	横向	0.055	0.071	0.855	0.223
	CRH380AJ-0201	垂向	0.116	0.148	0.739	0.434
	CRH380AJ-0201	横向	0.045	0.059	0.893	0.139
	CRH380BJ-0301	垂向	0.070	0.091	0.682	0.339
	CRH380BJ-0301	横向	0.060	0.081	0.780	0.104
GRU	CRH2A-2010	垂向	0.065	0.086	0.317	0.820
	CRH2A-2010	横向	0.041	0.056	0.396	0.711
	CRH2C-2150	垂向	0.083	0.106	0.366	0.753
	CRH2C-2150	横向	0.045	0.061	0.485	0.600
	CRH380BJ-A-0504	垂向	0.055	0.070	0.413	0.670
	CRH380BJ-A-0504	横向	0.037	0.048	0.346	0.782
	CRH5J-0501	垂向	0.041	0.052	0.413	0.701
	CRH5J-0501	横向	0.054	0.071	0.620	0.333
	CRH380AJ-0201	垂向	0.069	0.088	0.296	0.841
	CRH380AJ-0201	横向	0.043	0.056	0.611	0.383
	CRH380BJ-0301	垂向	0.056	0.074	0.420	0.667
	CRH380BJ-0301	横向	0.053	0.071	0.543	0.491
CNN-GRU	CRH2A-2010	垂向	0.050	0.065	0.227	0.902
	CRH2A-2010	横向	0.044	0.059	0.426	0.671
	CRH2C-2150	垂向	0.079	0.100	0.337	0.789
	CRH2C-2150	横向	0.046	0.061	0.443	0.621
	CRH380BJ-A-0504	垂向	0.054	0.068	0.408	0.691
	CRH380BJ-A-0504	横向	0.037	0.048	0.350	0.784
	CRH5J-0501	垂向	0.041	0.052	0.391	0.714
	CRH5J-0501	横向	0.060	0.078	0.631	0.243
	CRH380AJ-0201	垂向	0.058	0.074	0.245	0.888
	CRH380AJ-0201	横向	0.041	0.053	0.547	0.479
	CRH380BJ-0301	垂向	0.056	0.073	0.398	0.693
	CRH380BJ-0301	横向	0.062	0.080	0.545	0.416

下载: 导出CSV

表 5 各车型预测的加速度最大幅度值及偏差等级

Table 5. Maximum amplitude value and deviation level of the acceleration predicted by each track inspection vehicle

车型	垂向加速度/ （m·s⁻²）		横向加速度/ （m·s⁻²）
车型	最大值	超限等级	最大值	超限等级
CRH2A-2150	1.12	Ⅰ	0.57	未超限
CRH2C-2010	1.92	Ⅱ	0.53	未超限
CRH380BJ-A-0504	1.22	Ⅰ	0.30	未超限
CRH5J-0501	0.73	未超限	0.59	未超限
CRH380AJ-0201	1.58	Ⅱ	0.40	未超限
CRH380BJ-0301	0.98	未超限	0.84	Ⅰ

下载: 导出CSV

参考文献(21)

[1]	刘金朝,刘秀波. 轨道质量状态评价方法[J]. 铁路技术创新,2012(1): 106-109. doi: 10.19550/j.issn.1672-061x.2012.01.030
[2]	关庆华,赵鑫,温泽峰,等. 基于Hertz接触理论的法向接触刚度计算方法[J]. 西南交通大学学报,2021,56(4): 883-890. GUAN Qinghua, ZHAO Xin, WEN Zefeng, et al. Calculation method of hertz normal contact stiffness[J]. Journal of Southwest Jiaotong University, 2021, 56(4): 883-890.
[3]	LUBER B. Railway track quality assessment method based on vehicle system identification[J]. e & i Elektrotechnik und Informationstechnik, 2009, 126(5): 180-185.
[4]	LUBER B, HAIGERMOSER A, GRABNER G. Track geometry evaluation method based on vehicle response prediction[J]. Vehicle System Dynamics, 2010, 48(S1): 157-173.
[5]	FURUKAWA A, YOSHIMURA A. A method to predict track geometry-induced vertical vehicle motion[J]. Quarterly Report of RTRI, 2004, 45(3): 142-148. doi: 10.2219/rtriqr.45.142
[6]	FURUKAWA A. A method to predict vertical vehicle motion caused by track irregularities[J]. Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 2017, 231(4): 431-443. doi: 10.1177/0954409716634394
[7]	LI M X D, BERGGREN E G, BERG M, et al. Assessing track geometry quality based on wavelength spectra and track-vehicle dynamic interaction[J]. Vehicle System Dynamics, 2008, 46(S1): 261-276.
[8]	BERGGREN E G, LI M X D, SPÄNNAR J. A new approach to the analysis and presentation of vertical track geometry quality and rail roughness[J]. Wear, 2008, 265(9/10): 1488-1496.
[9]	LI M X D, BERGGREN E G, BERG M. Assessment of vertical track geometry quality based on simulations of dynamic track-vehicle interaction[J]. Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 2009, 223(2): 131-139. doi: 10.1243/09544097JRRT220
[10]	WANG J K, HE Y L, LU H Y, et al. Study on vibration acceleration prediction model of track inspection vehicle based on BP neural network[J]. IOP Conference Series: Materials Science and Engineering, 2018, 435: 012041.1-012041.8.
[11]	LI D, MEDDAH A, HASS K, et al. Relating track geometry to vehicle performance using neural network approach[J]. Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 2006, 220(3): 273-281. doi: 10.1243/09544097JRRT39
[12]	徐磊,陈宪麦. 轨道不平顺作用下铁路列车车体振动状态的PCA-SVM预测分析[J]. 铁道学报,2014,36(7): 16-23. doi: 10.3969/j.issn.1001-8360.2014.07.003 XU Lei, CHEN Xianmai. PCA-SVM forecast of car-body vibration states of railway locomotives and vehicles under the action of track irregularities[J]. Journal of the China Railway Society, 2014, 36(7): 16-23. doi: 10.3969/j.issn.1001-8360.2014.07.003
[13]	耿松,柴晓冬,郑树彬. 基于递阶神经网络的轨道车辆振动状态预测[J]. 城市轨道交通研究,2015,18(12): 94-98. doi: 10.16037/j.1007-869x.2015.12.021 GENG Song, CHAI Xiaodong, ZHENG Shubin. Prediction of rail transit vehicle vibration state based on hierarchical neural network[J]. Urban Mass Transit, 2015, 18(12): 94-98. doi: 10.16037/j.1007-869x.2015.12.021
[14]	李立明,柴晓冬,郑树彬. 基于径向基函数神经网络的轨道交通车辆振动状态预测[J]. 城市轨道交通研究,2017,20(12): 18-21. doi: 10.16037/j.1007-869x.2017.12.005 LI Liming, CHAI Xiaodong, ZHENG Shubin. Prediction of rail vehicle vibration state based on radial basis function nerve network[J]. Urban Mass Transit, 2017, 20(12): 18-21. doi: 10.16037/j.1007-869x.2017.12.005
[15]	刘秀波, 马帅, 高利民, 等. 基于轨道几何和LSTM的车辆响应预测模型[C]//第十三届全国振动理论及应用学术会议. 西安: 中国振动工程学会, 2019: 15-21.
[16]	MA S, GAO L, LIU X B, et al. Deep learning for track quality evaluation of high-speed railway based on vehicle-body vibration prediction[J]. IEEE Access, 2019, 7: 185099-185107. doi: 10.1109/ACCESS.2019.2960537
[17]	LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324. doi: 10.1109/5.726791
[18]	HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780. doi: 10.1162/neco.1997.9.8.1735
[19]	CHO K, VAN MERRIENBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha: Association for Computational Linguistics, 2014: 1724-1734
[20]	CHUNG J, GULCEHRE C, CHO K H, et al. Empirical evaluation of gated recurrent neural networks on sequence modeling[EB/OL]. （2014-12-11）[2021-11-10]. https://arxiv.org/abs/1412.3555.
[21]	康熊,刘秀波,李红艳,等. 高速铁路无砟轨道不平顺谱[J]. 中国科学(技术科学),2014,44(7): 687-696. doi: 10.1360/N092014-00088 KANG Xiong, LIU Xiubo, LI Hongyan, et al. PSD of ballastless track irregularities of high-speed railway[J]. Scientia Sinica (Technologica), 2014, 44(7): 687-696. doi: 10.1360/N092014-00088