Bearing Fault Fusion Diagnosis Method Based On Sparse Features Mining

In the rolling bearing fault diagnosis,due to the non-linear and non-stationary characteristics of vibration signals,the strong background noise in the working environment and signal transmission attenuation,there are problems such as difficulty in extracting fault features and insufficient ability to analyze features in a single domain.Fusion of multi-domain features extraction can provide more complementary fault information and improve fault diagnosis accuracy,but at the same time there are many problems that need to be solved,including how to mine potential sparse features based on bearing fault characteristics to achieve the purpose of complementing the information in the traditional analysis domain;how to mine the inherent discriminative information of the features;how to achieve effective fusion of different domain features.In view of these problems and the investigation on sparse essence of the rolling bearing fault signals,this thesis proposes a fault diagnosis method combining multi-domain features of rolling bearing based on sparse features mining,conducts in-depth research and proposes improvements in three aspects: features extraction,classifiers construction,and classifiers fusion,the purpose of which is to make full use of the discriminative information implied in the sparse representation of the features of each domain with complementary fault information,so as to improve the accuracy,reliability and robustness of the diagnostic system for rolling bearing fault diagnosis.The main research contents of this thesis are as follow:(1)For the pulse modulation phenomenon when the rolling bearing failure occurs and sparse characteristics of fault shock pulse,a cyclic pulse sparse features extraction method based on the fault band slice of S-transform is proposed.This method is based on the S-transform,determines the fault characteristic frequency based on the maximum impact energy of the fault and performs time-frequency slicing,and uses the cyclic window function to extract dimensionless pulse degree to extract the sparse features of cyclic stationary pulse fault,thereby achieving fault type diagnosis.The validity and superiority of the proposed method are verified and compared both in simulation signals and high-speed rail bearing fault signals in this thesis.(2)For different features have different fault sparse representation capabilities but sparse representation classifiers do not quantify feature weights with discriminative information,a feature-weighted sparse representation classification method based on iterative reconstruction sparse score is proposed.This method introduces the supervised sparse score of features,and obtains the optimal solution of feature weights through iterative reconstruction.Finally,the weighted feature set is used to construct a classifier.Experiments show that this method can effectively mine the potential discriminative information of features to improve the accuracy of sparse representation classification,and iterative reconstruction sparse score can largely overcome the problem of poor performance of traditional feature scoring algorithms in samples with multimodality and outliers.(3)For the decision problem in multi-classifier fusion diagnosis,a multi-sparse representation fusion fault diagnosis method based on confusion matrix estimated by reconstruction error probability distribution is proposed.This method uses estimates the reconstruction error in the sparse representation to mine the classification tendency of samples in each category,calculates an improved confusion matrix with better quantization performance for small samples,uses the improved confusion matrix to build a reliability matrix of classifier,and finally uses the combination rule to fuse recognition results from multiple classifiers.Experiments show that the improved confusion matrix measures classification discriminative information more accurately,and has a smaller sample number dependence,which can effectively improve the fusion recognition rate of multiple sparse representation classifiers under small samples.