基于改进图卷积神经网络的地铁转向架故障诊断方法

Fault Diagnosis Method for Metro Bogie Based on Improved Graph Convolutional Neural Network

摘要

为提高地铁列车转向架故障诊断的准确率与鲁棒性，提出一种融合卷积神经网络（CNN）与改进图卷积网络（GCN）的多传感器信息融合诊断方法。首先对各传感器采集的振动信号进行连续小波变换，生成时频图；其次利用5层CNN对各通道时频图进行特征提取，并将每个传感器视为图结构中的一个节点，提取的特征作为节点属性输入GCN；然后为克服传统GCN中邻接矩阵固定、无法反映节点间动态关联的缺陷，设计多层感知机（MLP）根据节点特征自适应更新邻接权重，实现动态图卷积；最后通过两层GCN与分类头完成故障识别。实验表明，所提方法平均诊断准确率达99.46%，在部分工况下最高可达100%，显著优于单通道CNN及CNN-LSTM、CNN-Transformer、MSSCNN等现有模型。混淆矩阵与t-SNE可视化结果显示，各类故障特征聚类清晰，错分率低于1%。研究证明，该方法在多工况、多故障模式下均具备优异的诊断精度与稳定性，为地铁转向架智能运维提供了可靠技术支撑。

Abstract

To improve the accuracy and robustness of fault diagnosis for metro train bogies, a multi-sensor information fusion diagnosis method combining convolutional neural network (CNN) and an improved graph convolutional network (GCN) is proposed. First, vibration signals collected from multiple sensors are transformed into time-frequency maps using continuous wavelet transform. Then, a five layers CNN is employed to extract features from each channel’s time-frequency map, where each sensor is treated as a node in a graph and the extracted features serve as node attributes for the GCN. To overcome the limitation of traditional GCN with fixed adjacency matrices that fail to reflect dynamic relationships among nodes, a multilayer perceptron (MLP) is designed to adaptively update adjacency weights based on node features, enabling dynamic graph convolution. Finally, fault classification is performed through two GCN layers followed by a classification head. Experiments demonstrate that the proposed method achieves an average diagnostic accuracy of 99.46%, and the maximum accuracy can reach 100% under some working conditions, significantly outperforming existing models such as single-channel CNN, CNN-LSTM, CNN-Transformer, MSSCNN. The confusion matrix and t-SNE visualization results show clearly clustered fault features with a misclassification rate below 1%. The study shows that the proposed method maintains excellent diagnostic accuracy and stability under multiple operating conditions and fault modes, providing reliable technical support for the intelligent operation and maintenance of metro bogie systems.

关键词

转向架 / 故障诊断 / 卷积神经网络 / 图神经网络

Key words

bogies / fault diagnosis / convolutional neural networks / graph neural networks

参考文献

[1] QIN N, DU J H, ZHANG Y M, et al. Fault diagnosis of multi—railway high—speed train bogies by improved federated learning[J]. IEEE Transactions on Vehicular Technology, 2023, 72(6): 7184-7194.

[2] GUO L R, LI G S, CHEN C, et al. Vibration fatigue characteristics of a high—speed train bogie and traction motor based on field measurement and spectrum synthesis[J]. Machines, 2025, 13(7): 613.

[3] MEN Z H, HU C Q, LI Y H, et al. A hybrid intelligent gearbox fault diagnosis method based on EWCEEMD and whale optimization algorithm—optimized SVM[J]. International Journal of Structural Integrity, 2023, 14(2): 322-336.

[4] 李刚, 秦永峰, 齐金平 . 基于GASF多通道图像时序融合的高速列车横向减振器故障诊断[J]. 振动与冲击, 2025, 44(15): 144—152, 191.

[5] LI G, QIN Y F, QI J P . High—speed train transverse damper fault diagnosis based on 1D—2D—GASF—CNN—GRU—MSA model[J]. Journal of Vibration and Shock, 2025, 44(15): 144—152, 191.

[6] 岑潮宇, 代亮成, 池茂儒, 等 . 基于KPCA—SO—KELM的抗蛇行减振器故障诊断[J]. 科学技术与工程, 2025, 25(11): 4551-4558.

[7] CEN C Y, DAI L C, CHI M R, et al. Fault diagnosis of yaw damper based on KPCA—SO—KELM[J]. Science Technology and Engineering, 2025, 25(11): 4551-4558.

[8] 门志辉, 宫岛, 周劲松, 等 . 基于格拉姆角场与并行卷积Transformer的列车传动系统故障诊断方法[J]. 机械工程学报, 2025, 61(24): 168-179.

[9] MEN Z H, GAO D, ZHOU J S, et al. Fault diagnosis method for train traction transmission system based on gram angle field and parallel convolutional transformer[J]. Journal of Mechanical Engineering, 2025, 61(24): 168-179.

[10] MEN Z H, LI Y H, GAO L, et al. Fault diagnosis method for railway wagon bearings under imbalanced dataset based on improved ACWGAN[J]. Nonlinear Dynamics, 2025, 113(12): 14935-14962.

[11] 彭刘禹, 胡俊锋, 张龙 . 基于GRCMFDE与CNN的轮对轴承故障诊断方法研究[J]. 铁道科学与工程学报, 2025, 22(9): 4260-4270.

[12] PENG L Y, HU J F, ZHANG L . Fault diagnosis of wheelset bearings based on GRCMFDE and CNN[J]. Journal of Railway Science and Engineering, 2025, 22(9): 4260-4270.

[13] MEN Z H, GONG D, ZHOU K, et al. Unsupervised domain adaptation method for bearing fault diagnosis assisted by twin data under extreme sample scarcity[J]. Mechanical Systems and Signal Processing, 2025, 239: 113359.

[14] 刘庆杰, 黄辉, 雷晓燕 . 基于集合经验模态分解和小波变换的轮轨力应变信号降噪[J]. 城市轨道交通研究, 2016, 19(11): 26—29, 37.

[15] LIU Q J, HUANG H, LEI X Y . Wheel/rail strain signals de—noising by using EEMD and wavelet transform[J]. Urban Mass Transit, 2016, 19(11): 26-29, 37.

[16] MEN Z H, CHEN Z, LI Y H, et al. Railway wagon bearing fault diagnosis method based on improved sparrow search algorithm optimizing variational mode decomposition and multi—level convolutional neural network[J]. Review of Scientific Instruments, 2024, 95(4): 045104.

[17] LI T F, ZHAO Z B, SUN C, et al. Multireceptive field graph convolutional networks for machine fault diagnosis[J]. IEEE Transactions on Industrial Electronics, 2021, 68(12): 12739-12749.

[18] QIN Y, WANG Y R, LI Z S, et al. An in—depth tutorial on BJTU—RAO bogie datasets for fault diagnosis[J]. IEEE Access, 2025, 13: 60879-60888.

[19] LI Y H, MEN Z H, BAI X N, et al. A bearing fault diagnosis method based on M—SSCNN and M—LR attention mechanism[J]. Structural Health Monitoring, 2025, 24(2): 830-852.

[20] MAATEN L V D, HINTON G . Visualizing Data using t—SNE[J]. Journal of Machine Learning Research, 2008, 9: 2579-2605.

基金资助

国家自然科学基金项目(52375115)

上海东方英才计划项目(QNKJ2025010)