Rolling Bearing Fault Diagnosis Based On Compound Associative Feature Optimization Model

Bearing has been widely applied in industry,which has been used in varied kinds of instruments,but also the most easily failed assemble unit.It is necessary to monitor the bearing's condition,because when it was failed,it will lead to the whole equipment broken down.When incipient fault first come out,the energy is always relatively small and easily mixed with environmental noise.So the energy should be amplified before diagnose the failure type occurred on the bearing.Maximum correlated kurtosis deconvolution developed from minimum entropy deconvolution,which has been proven as an effective filter for enhancing the impulses occurred in the rolling element bearing.The features of the signal should be extracted after filtering.Recurrence quantification analysis has advantage in extracting features from nonlinear and non-stationary signal,so this paper used it for feature extraction.Fault diagnosis of rolling element bearing is a kind of model recognition.Variable predictive model based class discriminate is a relatively new method which considers that different features have relationship with each other in nature,and these associations performance differently with different models.These models can be nonlinear and high order,so the VPMCD can be used as SVM and ANN in rolling element bearing fault diagnosis.In this paper,MCKD,RQA and VPMCD are used in rolling element bearing fault pattern recognition.The main content of this paper are as follows:(1)Based on the theory of MCKD method,the adaptive MCKD method combined MCKD with APSO method.The result shows that without a priori knowledge,AMCKD can enhance the fault signal.(2)Based on the theory of RQA,to classify the 11 kinds of feature,this paper presents an RQA-AP method combining Affinity propagation and RQA to extract and classify the fault signal features.(3)Based on the basic theory of VPMCD,a new multi-variable pattern recognition method combining APSO with VPMCD is proposed to solve the problem that VPMCD can only select one model when training.The new method adaptively fuse four model together to form the best one to identify the fault.According to the number of features of RQA is large,the features are clustered twice by AP,and four kinds of similar features and four different types are obtained and trained and identified by APSO-VPMCD.The results compared with SVM and VPMCD.The result shows that the difference of the feature will directly affect the accuracy of algorithm,and also shows the advantages of the new method.(4)In order to improve the classification accuracy of APSO-VPMCD at small samples,it is proposed to increase the training characteristics of APSO-VPMCD,that is,to train and identify with five kinds of features,and to reduce the time of AP clustering.The results shows that the new method has better stability and higher accuracy.According to the excellent performance of SVM,the Gaussian model is added to the VPMCD,the five models are predicted simultaneously.The experimental results show that the improved algorithm improves the accuracy and stability.