Health Assessment And Fault Diagnosis Of Train Communication Network Oriented To Smart Maintenance

As a critical system of rail train,train communication network(TCN)is an integrated central platform that is used to realize the train operation control,condition monitoring,and data transmission,whose failure will endanger the security of rail train running.The current maintenance mode of TCN is still corrective maintenance and planned maintenance,the lack of effective health assessment and intelligent diagnosis method is the fundamental reason for lagging maintenance timing,low efficiency and high costs.The application of prognostic and health management(PHM)technology can carry out state perception,fault diagnosis and prognostic of equipment or the system,which is an important means to change the traditional maintenance mode.The application of PHM technology in the maintenance of TCN is of great significance to improve TCN reliability and realize intelligent maintenance.To meet the demand of the intelligent maintenance of TCN,based on the systematical analysis of TCN fault mechanism,the overall architecture of health management is studied,the dissertation focuses on the key methods,including health assessment,anomaly detection and intelligent diagnosis,and solves especially the key problems in practical applications.The main contents of this dissertation are as follows:Firstly,for aiming at the problems existing in the current methods,a novel health assessment method for multifunction vehicle bus(MVB)network based on variational autoencoder(VAE)is proposed,the physical-layer signals are directly used as the input of VAE model to effective monitor the performance degradation of network communication.Compared with the deterministic mapping obtained by the traditional methods,the continuous probability distribution of normal samples is modeled by the VAE model,the health indicator of network node is obtained according to the average sample reconstruction error,the health weight of network nodes can be determined according to network topology characteristic index and network business requirement index,and the total health indicator of MVB network is calculated.The experimental results show that the proposed method has better robustness to noise,and has capable of quantifying the health condition of the network nodes and entire network,which can provide the theoretical support for condition-based maintenance.Secondly,to solve the problem of incomplete priori knowledge,an anomaly detection method for MVB network is proposed based on dynamic selective ensemble system,which involves two-class classifiers and one-class classifiers,and six waveform quality features are extracted from the physical-layer signals.The source competence of base classifier is estimated based on the information entropy and normalized gaussian potential function,and the competence classifiers are selected and anomaly detection result is obtained by weighted majority voting where the weights are equal to the competence values.The experiment results demonstrate that the proposed method outperforms other state-of-the-art methods without complete priori knowledge.Thirdly,considering the expensive and laborious labeling of the physical-layer signals,an intelligent fault diagnosis method for MVB network based on active learning(AL)and deep neural network(DNN),which is capable of building a competitive classifier with a limited amount of labeled training samples.To minimize the information loss,stacked consistent autoencoder(SCAE)is used to learn feature presentations from the physicallayer signals of master frame,which can improve the classification performance of DNN.In the supervised fine-tuning phase,a framework of the deep AL based fault diagnosis is proposed,and a dynamic fusion AL method is presented,the most valuable unlabeled samples are selected for labeling and training by considering uncertainty and similarity simultaneously,the fusion weight is dynamically adjusted at the different training stages.The experiment results show that to achieve the same diagnosis accuracy,the proposed method requires significantly less labeled training samples compared with the random sampling method.Also,under the same number of labeled samples,the proposed method achieve the better performance than state-of-the-art methods.Finally,aiming at the problem that it is difficult to learn the discriminative and highlevel features,a novel incipient fault diagnosis method for function module of train network equipment is proposed based on wavelet transform(WT)and multi-scale compactness convolutional neural network(MSC-CNN).Wavelet transform is utilized to decompose the raw signals into multi-scale components according to time and frequency information,a MSC-CNN architecture is developed,which can perform the multi-channel signals fusion and obtain complementary and rich diagnosis information from multi-scale components.With the joint supervision of the softmax loss and center loss,resulting in simultaneous maximization of interclass separability and intraclass compactness of samples.The experimental results demonstrate that the proposed method is very effective in feature extraction,and has higher diagnosis accuracy than other typical methods.