Research On High-Speed Train Bogie Fault Data Feature Analysis Method Based On Information Entropy Measurement

With development of high-speed railway in China, more and more people focus on security and stability of high-speed train. Bogie is the only connect unit between train body and rails. The performance degeneration and fault diagnosis of the key element of bogie will change the bogie vibration types, which seriously threat the safe operation of the train. Vibration signals of train body and bogie contain wealth of information. It is of great theoretical significance and practical value to achieve fault diagnosis, performance degenerate estimation and fault early warning for guaranteeing the safe and stability operation. However, the train vibration signal is typical high complexity, uncertainty and nonlinear. The traditional single feature extraction methods are unable to recognize fault exactly. It is necessary to find the novel feature extraction and fusion methods to achieve the bogie fault diagnosis and performance estimation.In view of this, the thesis proposes the feature extraction and analysis frame, which combines the information measurement entropy theory with time-frequency analysis method based on analyzing the physical meaning of the main indicatives of information measurement theory. According to the issues such as feature extraction of bogie fault, performance degeneration estimation of the key element and multiple feature fusion and dimension reduction, the main work of the thesis are as follows.1) Five kinds of wavelet imformation entropy are defined in the thesis. Wavelet imformation entropy’s physical meaning and the feasibility are researched for high-speed bogie fault. High dimension feature vector combined by multiple wavelet imformation entropies is used for failure state recognition of high-speed bogie key unit.2) A series of empirical mode decomposition entropies and complexity algorithms are proposed by combining information measure theory and ensemble empirical mode decomposition. The method is used for feature extraction of high-speed train bogie fault signal. At first, the simulation vibration signal in fault is decomposed by ensemble empirical mode decomposition. Then, a set of intrinsic mode functions of the decomposed results are selected. At last, the information measurement is calculated as feature. The fault type is identified correctly, so the features based on EEMD and information measure are effective.3) A new feature extraction method is proposed based on associated information measure. It is used for state estimation of bogie unit’s performance degradation. Incidence relation between vibration signal of bogie unit performance degradation and vibration signal of normal state is researched in the thesis. The incidence relation’s quantized value reflects the degree of performance degradation. Cross correlation sample entropy and relative ensemble empirical mode energy entropy are proposed for feature extraction of bogie unit performance degradation. The result of variance analysis and multiple analysis proves that the correlation information measure shows significant difference in each performance degradation stage. The probability density function of feature mean value is used to estimate the degree of performance degradation.4) For getting rid of bad features, increasing computational efficiency and reducing the feature dimension, multiple-information measure model based on multiple criteria feature selection and manifold learning is proposed. In the model, a high dimension feature set is built based on time domain, frequency domain, time-frequency domain and information measurement extended features. Relief algorithm, Mahamanobis distance and Fisher Ratio are used for feature ordering. The result of feature selection come from three orderings’weighted average. At last, manifold learning is used for feature dimension reduction. The experimental results of simulation data and measured data prove that Mulriple-information measure model is effective. The feature dimension is reduced, at the same time the recognition rate is significantly improved for all sensor signals.Plenty of simulation data experiment and measured data experiment results proved that the model and the feature extraction method are effective. A new research approach is proposed for bogie fault recognition and bogie unit performance degradation by data driven method.The research is supported by the Key Project of Natural Science Foundation of China-key problem research on high-speed train safety assessment based on data (No.61134002,2012-2016)...