Characteristic Analysis Of Vibration Signal Of High Speed Train Based On Ceemd And Feature Fusion

High speed train plays a very important role in our life which can hardly do without the development of high speed train technology. However, it is a two-edged sword and may bring the risk of safety. Therefore, it is necessary to research the safety of train. In view of the characteristics of nonlinear and nonstationary vibration signal of high speed train, the traditional signal analysis method has certain limitations, but complementary ensemble empirical mode decomposition (CEEMD) is suitable for analysis of vibration signal for its completeness and good adaptability. Further more, the single feature because of the fuzziness and uncertainty is difficult to achieve effective identification. Therefore, an effective feature extraction method is needed to solve the problem of safety state evaluation. This thesis put forward the characteristic analysis of vibration signal of high-speed train based on CEEMD and feature fusion.The main research contents in this thesis are summarized as follows:(1) Combined with CEEMD and information entropy, feature fusion is proposed for the characteristic analysis of vibration signal of high-speed train because the single feature has the characteristics of fuzziness and uncertainty. Firstly, by CEEMD, the complex signal is decomposed into a series of simple composition approximate stationary signal. Through the calculation of correlation coefficients, the intrinsic mode functions (IMFs) are chosed that has the largest correlation coefficients with the original signal. Then, the entropies of IMFs in time domain, frequency domain and in time-frequency domain are calculated to form a feature vector. Finally, feature vector is put into the least squares support vector machine (LSSVM) for classification and recognition. The results show that feature fusion has better performance compared with a single feature.(2) In view of different sensors to collect the information of both complementary and redundancy, this thesis proposes a multi-sensor feature selection method based on ReliefF. By sorting different sensor features weight using ReliefF, the sensor features are selected. Combined with the complexity measurement, multi-sensor feature fusion is proposed. We form a feature vector by calculating the weight and the fuzzy entropy of all selected IMFs instead of the mean of fuzzy entropy of all IMFs. It not only reduces the computational burden and improves the stability of the algorithm.The analysis of vibration signals on simulation experiment platform show that feature fusion has better recognition result.