Fault Diagnosis And Health Management For Train Communication Network

With the increasing intelligence of the train vehicles,more and more data are transmitted by the train communication network.Higher status monitoring and fault diagnosis requirements are put forward by the railway maintenance department.However,the current maintenance of the train communication network is still limited to the post-fault maintenance and planned maintenance.Lack of comprehensive network management capabilities makes it difficult to evaluate the health condition of the network and maintain the network before serious degradation.The means of network maintenance heavily depend on the engineer's expert experience,making the fault diagnosis and location difficult and inefficient.The final solution has always to be the massive replacement for equipment and cables.The disadvantage is not only the waste of maintenance resources but also the remaining of the root fault.In addition,the existing maintenance method is also incapable of dealing with intermittent and occasional faults,leaving the further deterioration of the fault and the hidden risk of the train.Therefore,it is of great significance to study the condition monitoring and health management of the train communication network.In this dissertation,we have extracted the network features of physical waveforms to characterize the network operating status.Based on the dataset and machine learning algorithms,the health evaluation,fault diagnosis,intermittent fault location of the network have been studied and discussed.The main work completed and the innovative results of the dissertation are as follows:Firstly,a train communication network PHM system framework and a feature extraction method based on the network physical layer waveform indexes have been proposed.In the PHM system,the network health evaluation,resistance mismatch fault diagnosis,intermittent connection fault location is completed.It lays a foundation for the intelligent operation and maintenance of the train communication network.Based on the dual-port model and analytic RLCG model,MVB in typical fault conditions is modeled and the transmission characteristic is obtained.According to the amplitude-frequency and phase-frequency characteristics of the transfer function,the causes of waveform distortion under network faults are analyzed.Based on the MVB analyzer,the high-speed A/D sampling of the network physical layer waveform signal is carried out.The state characteristics of MVB are extracted by means of the numerical fitting.Secondly,a network health evaluation method based on Support Vector Description Domain(SVDD)and sample reduction have been introduced.Aiming at the disadvantages of traditional health evaluation methods that rely too much on the subjective experience of experts,a classifier model representing the normal operating condition of the MVB network is trained by SVDD.According to the distance of the measured sample deviating from the hyper-sphere,the abnormal degree and health condition of the current network is quantified objectively and the health evaluation score of the whole network is obtained.The network maintenance plan is made according to it to lay the foundation of the condition-based maintenance of the train communication network.In the training process of the SVDD model,Density Based Spatial Clustering of Application with Noise(DBSCAN)is applied to realize the sample reduction operation and accelerate the training of the classifier.Thirdly,an impedance mismatch fault diagnosis based on Weighted Support Vector Machine(WSVM)has been proposed.The fault diagnosis proplem has been truned into a pattern recognition problem.To improve the accuracy of the classifier furtherly,a sample weight calculation method named multi-hop edge approach method is proposed.To validate the effectiveness of the proposed method,the algorithm has been conducted on the artificial synthetic datasets,and MVB datasets.Fourthly,an IC(intermittent connection)fault location method of MVB has been introduced.An IC frames recognition method based on device fingerprint and a fault location method based on network topology modeling are proposed.When IC fault occurs,the data frames are broken and difficult to decode.Once the configuration of MVB is inaccessible,the disturbed node is difficult to locate.Thus the concept of device fingerprint is introduced.The feature learning of normal data frames is carried out through the multi-layer sparse auto-encoder network.Then a three-layer fully connected neural network is cascaded to the auto-encoder and forms a deep neural network to characterize the network nodes' device fingerprints.When the IC fault occurs,IC frames are extracted by the MVB analyzer and the source nodes of those IC frames are identified based on the aforementioned deep neural network.The IC fault location is inferred based on network topology.Lastly,based on the summary of the whole research content,the dissertation gives the considerations and conclusions in the study process and proposes some problems which need further research in the future.