Research On Fault Diagnosis For Suspension System Of The High-speed Train In Complex Data Situations

The health monitoring of high-speed trains needs to handle complex data situations.On the one hand,an electric multiple unit train is a complex multi-body mechanical system.The operating condition of vehicles is affected by many external factors,including track irregularity,airflow,and load.There exist nonlinearity,strong coupling,and uncertainty in monitoring signals.On the other hand,the high-speed train contains many components with diverse characteristics,which requires identifying various fault patterns under an incomplete dataset.This thesis attempts to address the shortcomings of deep neural networks in the fault diagnosis of high-speed trains,aiming at improving the accuracy,robustness,and ability to handle incomplete data.The main contributions of this thesis are as follows.(1)A multi-perspective architecture for fault detection of high-speed train.The monitoring data of high-speed trains are collected from multiple sensors,which are both multi-scale and multi-component in nature.In order to handle these complexities,this thesis proposes a multi-perspective architecture of the neural network based on the combination of multi-branch structure and signal decomposition method,which is scalable and with low input dimensions.The proposed method provides multiple perspectives for the multi-channel and multi-component signal analysis,including perspectives from channel,component,and time scale.The proposed scheme improves the accuracy of fault diagnosis and has superior noise robustness,which could be valuable for practical applications of complex transportation systems,especially in dynamic environments.(2)A synchrony-based group convolution structure for multi-sensor fusion.A major technical challenge for a multi-channel monitoring system is the information fusion of different sensors with distinct types and divergence patterns.Thus,this thesis proposes a novel group convolutional network for the multi-sensor fusion based on synchrony information.The synchronization between multi-channel signals is quantified as the similarity indicator for the channel arrangement in the convolutional layer.In this approach,the features of signals with divergence patterns can be extracted more effectively.The proposed model outperforms normal convolutions and regular group convolutions on the fault detection task and has good computational efficiency as a result of model size reduction,which provides a feasible solution for the condition monitoring of the multi-channel monitoring system.(3)A out-of-distribution detection method for unexpected faults of high-speed trains based on Bayesian deep learning.The misidentification of infrequent faults could lead to unpredictable consequences for the vehicle’s safety.This thesis proposes a novel method for detecting unexpected faults based on Bayesian deep learning.First,the adversarial learning strategy is adopted to impose perturbation on input samples.Then,the difference between unexpected faults and known ones is magnified through a Monte Carlo-based uncertainty measure in the Bayesian deep learning scheme.Finally,the uncertainty is qualified with mutual Information,which is used for determining whether the anomalies belong to known classes.The proposed scheme has a solid theoretical foundation and can be implemented in the fine-tuning model,which means it can achieve detection with no extra training for the model.Besides,only a few samples of unexpected anomalies are required for calibration,which is feasible in practical scenarios.(4)A recognition method for simultaneous-fault of high-speed trains based on betweenclass learning.The simultaneous fault can be difficult to detect accurately,owing to the tightly coupled components and incomplete data.This thesis proposes a novel method for simultaneous-fault diagnosis based on between-class learning.First,the neural network is trained with the multilabel approach using between-class samples synthesized based on channel energy.Next,the enhanced estimation based on Bayesian deep learning is designed to amplify the discrimination score of suspicious samples of simultaneous faults.The proposed method can achieve both single-fault and simultaneous-fault detection in one model and identify corresponding components of simultaneous faults.In addition,the extra cost of training data collection can be greatly reduced because no real multi-label sample is required in the training phase.