Research On Bogie Fault Diagnosis Based On Sub-Signal Trend Analysis And Fixed Dictionary Extreme Learning Machine

With the acceleration of China’s urbanization process and the rise of TOD(transit-oriented development)urban development model,the safety and stability of metro which is the main transportation system,are closely related to people’s production and life and the economic development of the country.Ensuring the safety and stability of metro operation is of great significance to maintain the development of the country and society.As an important part of the metro health management system,fault diagnosis can evaluate the health status of metro and guide the related maintenance and repair work.The use of it is an indispensable and important means to ensure the efficient and stable operation of metro.In the field of metro fault diagnosis,the bogie has always been the primary research object.As an important force bearing part of metro,bogie is one of the most vulnerable parts to be worn and failure,and its running state directly affects the safety of metro.Therefore,the fault diagnosis of metro bogies can well maintain the safe and stable operation of subway trains,and it is of great significance to protect the safety of people’s lives and property and improve the operation efficiency of economic society.Based on it,in this paper,the bogie of Guangzhou Metro A2 Bullet Train is taken as the specific object to carry on the research of fault diagnosis.Finally,the bogie fault diagnosis method based on sub-signal trend analysis and fixed dictionary extreme learning machine is formed.The specific research process is summarized as follows:(1)This paper firstly analyzes the influence of the variable speed of metro and the low SNR environment on the vibration signal waveform,and simulates it with MATLAB.It is found in the simulation that the influence of variable speed and low signal-to-noise ratio will greatly change the time-frequency domain characteristics of the signal.This change will reduce the efficiency of the existing time-frequency signal analysis methods.Moreover,the improvement of noise reduction and waveform decomposition carried out by many scholars in these methods will get worse because of the change of waveform characteristics,and it also will cause a lot of waste of computing power.(2)In order to solve this problem,the singular value decomposition is firstly proposed in this paper to improve the extraction effect of singular value decomposition in weak fault features of low SNR signals.In the study of improvement,this paper analyzes the recently discovered formulas in the field of physics which describe the intrinsic relations between the elements in the eigenmatrix.This formula has been extended to more general cases by spectral theorem and can be used in normal matrices.In this paper,the singular value decomposition is further derived according to the formula,so that the feature information hidden in the left and right singular matrices is effectively utilized,and the fully decomposed singular value algorithm is formed for feature extraction.(3)After that,in order to solve the problem of waveform distortion caused by the variable speed and the limitation of sensors sampling frequency,this paper proposes a sub-signal trend analysis algorithm based on the fully decomposed singular value algorithm.The algorithm firstly decomposes the signal into two parts,high and low frequency respectively.And then it uses the upper and lower envelope fitting method to restore the waveform trend.After that,in the processed waveform,the periodic waveform which is consistent with the change of mechanical rotation period is found for analysis,and the fully decomposed singular value algorithm is used to complete the feature reconstruction.This algorithm has a strong ability to resist waveform distortion,and further enhances the ability to resist noise.The running speed of it is also optimized.In the signal with SNR lower than 10 d B,the algorithm can obtain excellent features with rapid speed.Moreover,in the classification test based on extreme learning machine,the characteristics obtained by variational mode decomposition,empirical mode decomposition,ensemble empirical mode decomposition and local mean decomposition which were all combined with singular value decomposition respectively were lower than sub-signal trend analysis algorithm by 16%-53% in maximum identification accuracy.(4)Finally,in the part of pattern recognition,in order to improve the training stability of the Extreme Learning Machine,meeting the robustness requirements of metro bogie fault diagnosis,and solving the problem of large fluctuation of recognition accuracy caused by the local optimization of the Extreme Learning Machine.Based on the thoughts of sparse coding in hierarchical extreme learning machine and the dictionary learning of K-SVD,this paper designs a new automatic coding method and proposed the fixed dictionary extreme learning machine.This algorithm cancels the bias matrix,and assign the weight matrix by the empirical formula,which greatly improves the stability of the algorithm training.Compared with the original algorithm,the training speed and recognition accuracy of the proposed algorithm are improved obviously.In the comparison experiment with original extreme learning machine,hierarchical extreme learning machine and support vector machine,the algorithm can obtain the recognition accuracy of 1.5%-12% higher than the comparison algorithm with extremely fast training speed,and it also can guarantee the stability of the training effect just like the support vector machine.