Research On Fuzzy KNN Based On Dempster-shafer Theory And The Application In Fault Diagnosis

With the rapid development of modern economy, the modern industry’s scale isbecoming lager and the degree of the complexity of industrial process is higher andhigher, and as a result, the modern system is more and more complex. Fault diagnosismethods based on data driven are studied in this paper especially the machine learningmethods given the situation that the accurate model of a complex system is very hardto obtain while the historical data of the system under the conditions of normal and allcategories of faults are relatively easy to get.kNN is one of the machine learning methods which is commonly used. As atheoretical perfect method, kNN has the advantages of simple and higher correctclassification rate while not requiring the process of training. From the above, kNNand its application in fault diagnosis are further improved in this paper based on theresearch on kNN and a variety of modified kNN methods. The main work andcontributions are as follows:(1) kNN and the modified kNN methods especially FkNN and EkNN are studiedparticularly in this paper, meanwhile, the advantages and the disadvantages of thesemethods are concluded.(2) A method called FEkNN is presented in this paper to overcome theshortcomings of kNN and the modified kNN methods especially FkNN and EkNNsummarized above. In this method, two problems that the differences of the samplefeatures can’t be recognized and the effect of fuzziness that aroused by the differentdistances between neighbors and the center of classes is not taken into account aresolved.(3) A fundamental problem in multi-classifier system is how to combine thedecisions made by the classifiers according to some way to reach consensus given theclassifiers’ different performances. Taken this into account, the classifiers’ decisionsneed to be weighted for a more accurate identification rate. The weights are normallydefined as the correct rates corresponding to classifiers calculated by the training samples, and then the final decision is made by means of weighted average. Aproblem arouses in the method just described that the weights are not accurate enoughget form the classifiers’ correct rates. A new method is presented in this paper for thisproblem in which the weights are figured up in another way.(4) The proposed methods are applied to fault diagnosis and its effectiveness isillustrated by UCI standard data set and electric actuator fault diagnostic experiment.