Research On State Prediction And Fault Diagnosis Of High-speed Train Bogie Based On Graph Neural Networks

With the continuous expansion of the scale of Chinese high-speed railway network,the number of high-speed railway trains has increased significantly,and the safety of high-speed railway trains in transit has become an important issue in current research.As a key subsystem of high-speed trains,the train bogie system is the only structure connecting the car body and the track.It can ensure the stability of the car body,and also play an important role in traction,braking and buffering.Since the high-speed train bogie system is a system composed of multiple mechanical equipment and electrical equipment,the long-term train operation will cause prob lems such as wear,aging,and short circuit of the equipment.In order to ensure that the bogie system is in a healthy state during train operation,it is also very meaningful to carry out research on the fault diagnosis and health state prediction of the bogie system.However,the bogie is greatly affected by the operating environment and working conditions during the operation,and the state and fault characteristics of the bogie are nonlinear.The traditional state prediction methods and fault diagnosis methods cannot meet the actual needs,and there are not many related researches on the multi-sensor of the high-speed train bogie system.Therefore,combined with the monitoring data of the high-speed train condition monitoring system,this dissertation uses the theory and method of deep neural network to construct a graph-based high-speed train bogie system state prediction model and fault diagnosis model.The main content of this dissertation includes the following six aspects:(1)Representation of train graph structure features based on GGC methodThe GGC method uses the sensor measurement points as the nodes of the graph structure and the correlation relationship between the measurement data as the edges of the graph structure.It uses Granger Causality to define the direction of graph structure edges.In this dissertation,the graph structure of the bogie system is given using the GGC method,and the bogie graph structure is proved to be spatially correlated using the Moran index.(2)Short-time prediction method of high-speed train bogie based on GCGThe GCG method uses the Gate Recurrent Unit to extract the temporal features of train bogies and the Graph Convolutional Network to extract the spatial features of train bogies.It replaces the original linear unit of GRU with a graph convolutional module to achieve state prediction.We experimentally demonstrate that the GCG model can effectively achieve the short-time prediction of train bogie state,and has better sensitivity and tracking ability for the abrupt change problem.(3)Long-term prediction method of high-speed train bogie based on GA-GRGATThe GA-GRGAT model uses the Graph Attention Network to extract the spatial features of the bogie directed graph.It embeds the GAT model into the GRU model to extract the temporal features of the bogie system,and uses the attention mechanism and Seq2 Seq architecture to reduce the cumulative error generated by multi-step prediction.It uses the Generative Adversarial Networks to construct time-conditional sequences to fuse the historical train bogie state information.The experimental results demonstrate that the accuracy of GA-GRGAT model is over 90% for long-time prediction(one day)and over 80% for ultra-long-time prediction(2 weeks),which is much better than other prediction models under the same conditions.(4)Undirected graph multi-sensor fault diagnosis method for high-speed train bogies based on Att GGCNThe Att GGCN model uses the GCG framework to fuse the spatio-temporal features of train bogies and further fuse the fault features using the attention mechanism to perform train fault diagnosis.It is demonstrated that the Att GGCN model can effectively identify seven bogie fault categories with high accuracy and robustness.(5)Directed graph multi-sensor fault diagnosis method for high-speed train bogies based on RS-GATThe RS-GAT model uses the Residual-Squeeze Network,replacing the convolutional layers in RS-Net with GAT operations.It uses 6-layer GAT residual units for deep fault feature extraction on train bogie data.The experimental results demonstrate that the fault recognition accuracy of RS-GAT model is higher than 90%,and the model is more interpretable.The research results obtained in this dissertation enrich the research contents in the field of health state prediction and fault diagnosis of high-speed trains at home and abroad in a certain sense,and can also provide theoretical and empirical support for managers and decision-makers of high-speed train operation control.