High-Speed Train Running Gear Fault Feature Extraction Based On Manifold-Learning And Compressive Sensing

Running gear fault signal, which was monitored by different acceleration sensors from different measuring points, was a linear or nonlinear mixture of varied vibration modes made by failure sources as well as other structure units of running gears, excited by track irregularities. Besides, given the limited measurement conditions, vibration signal was inevitably stacked with certain noise. Hence, it is very difficult to qualitatively analyze different fault types by using vehicle dynamics modeling, force analysis or system differential equations. Reversely, focusing on the feature analysis of monitoring data to recognize the effect that components failure had on high-speed train’s dynamics performance shows some superiority, given the truth that monitoring data surely has fault information of the damaged components. Based on three kinds of monitoring data:No.1, running gear fault simulation data(single faults, composite faults under 10 operating conditions included); No.2, Wuhan-Guangzhou PDL vehicle GPS measured data of wheel tread wears under 4 operating conditions; No.3, the standard data set of bearing faults(including single faults data under 4 operating conditions as well as inner-loop performance degradation data under 3 operating conditions); A new model of feature extraction and feature selection was proposed in this thesis.Firstly, to have a better understanding of the features and effects when there were some components faults happened to the running gear of High-Speed trains, wave analysis in both time-domain and frequency-domain, preliminary features such as 6-dimension statistical features and the 20-dimension singular value features were studied in this paper. On purpose of making sure if the preliminary features extracted before can identify different fault types effectively or reflect the deterioration process of components properly. What’s more, another experiment was conducted to make sure if the train operation stability was affected when running speed varies.Secondly, based on running gear fault simulation data of absorbers completely removed under 7 operating conditions, two methods:multi-scale analysis and information entropy theory, which is widely used in fault analysis fields, were employed to form original high-dimensional feature set of each sample. This feature set consists of statistical features, multi-scale entropies, and singular values. Then, the supervised ISOMAP algorithm(S-ISOMAP) of Manifold-Leaning was used for dimensionality reduction, low-dimensional optimal features were got. Finally, to further prove the validity and superiority of those low-dimensional features, Fisher-Ratio algorism was introduced to evaluate both the high-dimensional and low-dimensional features. The results indicated that S-ISOMAP method can removal the redundancy among high-dimensional features when most of the effective information was kept. The low-dimensional features which had a more preferable divisibility, were effective parameters in separating seven fault types.Thirdly, to recognize if the train operation stability changes when the wearing degree of wheel tread got deeper, singular spectrum relative entropy as well as grey absolute relational grade based feature analysis method were studied. The result showed, when locomotive wheel tread was unworn, the value of relative entropy approached to zero, while the grey relational grade value tended to one. As the degradation degree got deeper, the similarity between normal condition signal and heavy-degraded state signal was very small. In this case, the relative entropy value got larger, while the grey relational grade value got smaller.Finally, quadratic statistical feature extraction as well as compressive sensing(CS) based feature dimension reduction were hired to study the inherent law of each fault, The validity of the proposed method was verified by both the standard dataset experiment of bearing faults and the running gear fault dataset experiment. That is:1).quadratic statistical features analysis result was more preferable than the method of directly extracting statistical features from raw signals; 2).low-dimensional features which were got from CS based dimension reduction, achieved a good fault isolation.