Classifier Design Based On Evidence Reasoning Rule And Rough Set And Its Application In Fault Diagnosis

The Evidence Reasoning(ER)rule is an effective extension of the Dempster combination rule in Dempster-Shafer(DS)evidence theory.It solves the abnormal change problem of specificity before and after a piece of evidence is discounted.The concepts of importance weight and reliability factor of evidence are distinguished.Recently,some scholars have applied ER rule to equipment fault diagnosis and given certain results.On this basis,this paper gives a method to obtain the reliability factor by quantifying the uncertainty of the classification attribute sample and its evidence.Based on this,a new type of generalized classifier is designed,and then the classifier is applied to fault diagnosis under complete and incomplete samples,so as to improve the applicability and effect of ER-based diagnosis model in solving equipment fault diagnosis problems.The main contents are as follows:(1)A generalized classifier is designed by using evidence reasoning(ER)rule and rough set.Firstly,the reference evidence matrix is obtained by the sample feature reference value and the reference evidence matrix parameters are optimized by the evidence uncertainty.Then,based on the rough set and the uncertainty of the evidence to obtain the reliability factor of the evidence,ER rule is used for evidence fusion and classification decision.Finally,using the five typical classification data sets provided by the University of California Irvine(UCI),the implementation process of the generalized classifier is described in detail,and the experimental results are analyzed in detail to illustrate the advantages of the proposed classifier performance.(2)Fault diagnosis method for rotating machinery under complete sample conditions.For the case where the training sample is complete,that is,there is no missing in the multi-dimensional fault feature vector required for modeling,the corresponding diagnosis method is given based on the generalized classifier in(1).Firstly,the mechanical vibration signals collected by multi-sensors are extracted.The k-means algorithm is used to obtain the reference values of the feature samples,the reference evidence matrix is constructed and the evidence reliability factor is calculated.The evidence fusion and decision-making are used to judge the failure mode of the samples to be detected.The sequential linear programming(SLP)optimizes the importance weight of the evidence.Finally,the effectiveness of the proposed method is verified in the motor rotor fault diagnosis experiment.(3)Fault diagnosis method for rotating machinery under incomplete sample conditions.Aiming at the typical problem of missing sample data in fault diagnosis,two further diagnostic methods based on generalized classifier in(1)are proposed.In the first method,the missing fault features are ignored in the process of multi-source evidence acquisition and fusion;the second method considers the data loss as a natural characteristic of the fault feature itself,and considers it in the evidence acquisition and fusion.Finally,in the motor rotor fault diagnosis experiment,the effectiveness of the proposed two methods is verified by comparing other typical diagnostic methods.