| With the implementation of the national energy substitution strategy and environmental protection policy,the new energy vehicle industry has been developing under the corresponding policy incentives,among which electric vehicles are the main research object in the new energy vehicle field.As the key power source of electric vehicles,the drive motors need to satisfy the requirement of short-term overload,high temperature resistance,wide range of speed change and frequent start-stop.Rolling bearing is a key component of drive motor which is easy to be damaged.It is subjected to complex alternating stress and thermal stress.Long time running will lead to the inevitable rolling bearings failure,which will directly affect the running performance of the driving motor,and then lead to the decline of vehicle handling stability and ride comfort,and even endanger the driving safety of electric vehicles.Therefore,this thesis takes the electric vehicle drive motor rolling bearing as the research object,and carries out fault pattern recognition on its vibration signal,so as to provide the decision basis for the maintenance and life prediction of the drive motor according to the situation.Firstly,for bearing failure vibration signal with non-linear and strong noise interference problem,a fault signal denoising method based on adaptive variational mode decomposition was proposed.With the correlation waveform factor as the objective function and the amplitude spectrum of the envelope spectrum as the computing domain,the number of modes and penalty factor of the variational mode decomposition were selected by the Archimedean optimization algorithm,and the eigenmode components containing fault information were obtained.According to the characteristics of rolling bearing fault vibration signals,simulation signals were constructed,and the proposed adaptive variational mode decomposition method was used for noise reduction.By comparing with the decomposition results of other improved methods,the effectiveness and superiority of the proposed method were verified.Secondly,a fault entropy feature extraction method and stack sparse encoder was proposed to solve the problem that weak fault features in intrinsic mode component are difficult to identify.The envelope spectrum entropy,multi-scale permutation entropy,variancedelay fuzzy approximate entropy and refined composite multi-scale dispersion entropy of the first two eigenmode components were calculated to construct the fault feature data set.A stack sparse automatic encoder and Softmax classifier were used to reduce the dimension of the fault feature data set.Particle swarm optimization algorithm was used for selecting the hidden layer sparse constraint term and regularization term coefficients to obtain the optimal dimensionality reduction fault feature matrix.Finally,aiming at the problems of poor nonlinear data mapping ability of traditional machine learning methods and complex structure of deep neural network model,a shallow neural network fault pattern recognition method based on kernel extreme learning machine was proposed.Taking Case Western Reserve University fault bearing and the Xi ’an Jiaotong University rolling bearing life dataset as experimental objects,kernel extreme learning machine was used to establish the mapping relationship between fault characteristics and fault modes.The traditional machine learning model and the classical neural network model were compared with the proposed method.The model performance was verified via the improved Kfold cross-validation,and the accuracy rate,precision,recall rate and F1 score were verified. |
