High–speed railway track quality index prediction method based on 3D convolution neural network
-
摘要: 轨道质量指数(TQI,Track Quality Index)是反映高铁整体线路质量状态的重要指标,分析TQI数据的变化规律能够对高铁线路养护维修提供重要指导和参考依据。为提高TQI数据预测的准确性,提出了一种多项特征数据的3D卷积神经网络模型,分析了TQI数据特征,抽取时间、空间、检测项数据并形成三维特征数据集,基于3D卷积神经网络算法,构建8层TQI预测模型,并从初始化参数、学习速率、激活函数、损失函数、Dropout方法等角度对模型进行优化,并利用某高铁线检测数据进行试验验证。结果表明,3D卷积神经网络模型可较好的预测高铁线路状态变化趋势,且对比于BP神经网络和2D卷积神经网络方法,平均绝对误差分别降低了41.48%、26.32%,均方差分别降低了65.42%、39.93%,证明了该方法的准确性与有效性,对于预测TQI与制定高铁线路养护维修计划具有实用价值。Abstract: Track Quality Index(TQI) is an important indicator reflecting the overall quality status of the high-speed railway, thus the analysis of the change pattern of TQI data can provide important guidance and reference for the maintenance and repair of high-speed railway. This paper proposed a 3D convolutional neural network model with multiple feature data, analyzed the characteristics of TQI data, extracting time, space and detection items to form a three-dimensional feature data set. Based on 3D convolution neural network algorithm, the paper constructed an 8-layer TQI prediction model, and optimized the model from the aspects of initialization parameters, learning rate, activation function, loss function, Dropout method, and used the detection data of a high-speed railway line for experimental verification. The test results show that the 3D convolution neural network model can better predict the state change trend of high-speed railway lines. Compared with BP neural network and 2D convolution neural network, the average absolute error is reduced by 41.48% and 26.32%, and the mean square error is reduced by 65.42% and 39.93%, respectively. It is proved that the method is accurate and effective, and has practical value for predicting TQI and formulating maintenance plan of high-speed railway lines.
-
表 1 3D卷积神经网络结构
层数 核结构 卷积步长 激活函数 输出形状 输入层 / / / 5*5*7 3D卷积层1 32*(5*5*7) 1*1*1 Mish 5*5*7 3D池化层1 2*2*2 4*2*2 / 2*3*4 3D卷积层2 64*(2*4*6) 1*1*1 Mish 2*3*4 3D池化层2 2*2*2 4*2*2 / 1*2*2 2D卷积层1 128*(1*2*2) 1*1 Mish 1*2*2 2D池化层1 2*2 1*1 / 2*2 全连接层 / / Sigmoid 1 表 2 3类预测方法各指标列表
预测方法 评估指标 MAE MAPE MSE RMSE 3D卷积神经网络 0.182 5.44% 0.039 0.1979 LeNet-5卷积神经网络 0.247 7.39% 0.065 0.25 双隐层BP神经网络 0.31 9.31% 0.11 0.33 表 3 3类预测方法指标对比
方法对比 误差减少(%) MAE MAPE MSE RMSE 3D卷积神经网络相比于LeNet-5卷积神经网络 26.32 26.32 39.93 22.48 3D卷积神经网络相比于双隐层BP神经网络 41.48 41.48 65.42 41.19 -
[1] 汤国防. 铁路轨道几何不平顺变化特征及其预测模型研究[D]. 北京: 北京交通大学, 2010. [2] 张 念. 铁路轨道几何不平顺趋势预测的关键算法研究[D]. 成都: 西南交通大学, 2015. [3] 潘海泽,黄远春,汪 磊,等. 基于灰色GM(1, 1)和灰色-马尔可夫模型的轨道几何不平顺预测及应用研究 [J]. 铁道标准设计,2010(10):5-9. DOI: 10.3969/j.issn.1004-2954.2010.10.002 [4] 宋博洋. 基于时间序列模型的轨道质量指数预测研究 [J]. 山东科学,2014,27(3):66-72. DOI: 10.3976/j.issn.1002-4026.2014.03.013 [5] 韩 晋,杨 岳,陈 峰,等. 基于非等时距加权灰色模型与神经网络的轨道不平顺预测 [J]. 铁道学报,2014,36(1):81-87. [6] 冯 超,余朝刚,孙 雷,等. 基于改进GM(1, 1)与WOA-LSSVM组合预测模型的轨道不平顺预测 [J]. 铁道标准设计,2019,63(4):34-39. [7] 王建西,李海锋,许玉德. 基于概率分布推移变化的铁路轨道几何状态评价与预测方法 [J]. 中国铁道科学,2008(5):31-34. DOI: 10.3321/j.issn:1001-4632.2008.05.006 [8] Tanaka H, Yamamoto S, Oshima T, et al. Methods for Detecting and Predicting Localized Rapid Deterioration of Track Irregularity Based on Data Measured with High Frequency [J]. Quarterly Reports of RTRI, 2018, 59(3): 169-175. DOI: 10.2219/rtriqr.59.3_169
[9] Falamarzi A, Moridpour S, Hesami R, et al. Rail Degradation Prediction Models for Tram System: Melbourne Case Study [J]. Journal of Advanced Transportation, 2018(Pt. 2): 407-414.
[10] Guler H. Prediction of railway track geometry deterioration using artificial neural networks: a case study for Turkish state railways [J]. Structure and Infrastructure Engineering, 2014, 10(5): 614-626. DOI: 10.1080/15732479.2012.757791
[11] 彭丽宇,张进川,苟娟琼,等. 基于BP神经网络的铁路轨道几何不平顺预测方法 [J]. 铁道学报,2018,40(9):154-158. DOI: 10.3969/j.issn.1001-8360.2018.09.021 [12] 杨怀志,吴艳华,程智博. 基于铁路数据服务平台的高速铁路工务设备大数据应用研究 [J]. 铁路计算机应用,2018,27(7):93-97. [13] 张 颖,袁和金. 基于3D卷积神经网络的人体行为识别方法 [J]. 软件导刊,2017,16(11):9-11. -
期刊类型引用(4)
1. 李帅,王志飞,李樊,杜呈欣,王浩东,杨博璇. 基于机器视觉的高速铁路站车联动控制技术. 科学技术与工程. 2025(02): 773-779 . 百度学术
2. 鲁佩林,王凯. 基于3D卷积神经网络的深基坑施工邻近建筑物沉降预测. 建筑技术. 2024(S1): 95-97 . 百度学术
3. 张梓鸿,赵正阳,王文斌,李洋,李沅军. 基于卷积神经网络的城市轨道交通轨道几何不平顺指标预测. 城市轨道交通研究. 2023(10): 109-115 . 百度学术
4. 李强,刘铁生,崔霆锐,黄玉,张志亮. 基于神经网络的焊接接头不平顺故障诊断研究. 铁道车辆. 2022(06): 69-72 . 百度学术
其他类型引用(8)