Research On Evaluation Method Of High-Speed Train Bogie Fault Diagnosis

Parameter optimization and fault feature extraction about the key components of the bogies have been widely researched.However,the research of fault diagnosis method is not carried out.In practical,fault diagnosis results generally include some key information such as fault type,location,degree,and post-processing.In this paper,the fault decision-making process for high-speed train bogie is divided into five steps:fault detection,fault classification,fault location,fault quantification and post-processing.The secondary suspension system is discussed in this paper.Fault detection,fault classification and location are mainly discussed in following contents based on two aspects consist of feature selection or reduced dimension,classifier model training.1.According to the geometric symmetry structure of vehicle body,7 different representative single fault conditions and normal condition about air spring,lateral damper and yaw damper are selected to study.The 12 dimensional statistical features,8 dimensional wavelet energy moment features and 3 dimensional chaotic features are extracted from vibration signal's time domain,frequency domain and time-frequency domain.The 18 independent channels on the frame and the vehicle body are selected to constructed multi-channels feature space based on series connection,which describe the train working condition information from different aspects.Moreover,the multi-channels feature space has 414 dimensional.2.In order to detect whether the secondary suspension system is fault or not,the fault detection method is taken into account.Firstly,the Laplacian Eigenmap(LE)algorithm is utilized to reduce the dimension of the original dataset.Because the original data set belongs to two different manifold structures,the neighborhood graph need to be decomposed,and then it is processed by LE.Afterward,Low dimensional manifold data are detected by Support Vector Data Description with negative samples(SVDDne)model.The missing rate of LE+SVDD algorithm is 0%,and the accuracy of which is up to 99.31%.Algorithm LE+SVDD is outperform then other two methods via the comparison of missing rate,false rate and accuracy.A new idea for fault detection of high-speed train is put forward.3.In order to determine which kind and which position component is in fault condition,the fault classification and location methods are studied.Firstly,the hierarchical structure is constructed according to original features utilizing the stand K-means clustering method and the K-means based on genetic algorithm.The SVM algorithm is regarded as base classifier of hierarchical classification,Binary Decision Tree(BDT)structure is used to memory and decision direction adopts top-down strategy.Then,original features of the six nodes in hierarchical structure are selected with SVM-RFE and SVM-RFE with Correlation Bias Reduction(SVM-RFE+CBR)algorithm.Finally,the stability of feature selection algorithm is analyzed via fusion method.The dimension and accuracy of optimal feature subset of SVM-RFE+CBR are better than SVM-RFE's.The recognition rate is increased about 1.6%,about 5%higher than the original feature set.It is shown that the original feature space contains high correlation feature group.The stability of feature selection method is improved with the number of fusion increasing.