FCM Clustering Algorithm And Its Application In Fault Diagnosis Of Transformer

The state of power transformer affects the stability of the whole power system directly.Because the three-ratio method has unique advantage, it is used most widely in practiceamong many kinds of transformer fault diagnosis methods. But the corresponding relationshipcodes of it and fault type is too strict, some codes of it are lacking, so its fault diagnosisaccuracy rate is not high.Fuzzy clustering algorithms can divide samples softly in the way of me mbership,based on the fuzzy future of transformer fault diagnosis, this paper uses FCM (Fuzzyc-Means clustering) algorithm to judge transformer fault types co mbining with three-ratiomethod. Firstly, for the shortcomings that it does not distinguish the different effect ofevery sample, we improve the FCM algorithm, according the idea that Euclidean distancecan express the dissimilarity of samples, then we achieve the SWFCM(sa mple weightedFCM) algorithm. The experimental results based on IRIS data sets show that SWFCMalgorithm can get cluster center which is closer to the position of actual cluster center thanFCM algorithm.When we use SWFCM algorithm to diagnose types of transformer fault, we arrange thediagnosis data of standard samples and the sample to be tested in a matrix according the orderof types which exist in the three-ratio method, this matrix is the input of SWFCM algorithm.The algorithm can automatically identify the type of fault of the sample tested through thecomparison of membership degree. Instances show that this fault diagnosis model can makefault diagnosis accuracy to be93.42%, and the anti-interference ability of the algorithm isstronger than the three-ratio method after the initial data is added noise.In initialization, this paper sets up more reasonable initial value in membership degreematrix for SWFCM algorithm, according the actual model of diagnosis. From the number ofiterations, we can see SWFCM algorithm significantly improves convergence speed, it candetermine the fault type more quickly than FCM algorithm.To reflect SWFCM algorithm’s function on modifying cluster center position, we divide152samples using SWFCM algorithm directly, because the numbers of sample in every kindare very different, the kinds which has less samples will be treated as scattered or isolated.The error sum of squares of cluster center output by FCM algorithm is more large than theoutput of SWFCM algorithm obviously, this shows that the kinds which contains less sampleshas large negative impact for clustering, now the accuracy of fault diagnosis is quite low. Ifwe extract20samples respectively from three kinds which contain much more samples, anduse the two algorithms to cluster for the samples extracted, the difference between the error sum of squares of cluster center output by SWFCM algorithm and FCM algorithm is smallerthan before. This result indicates negative impact generated by scattered or isolated kindsdrops. SWFCM algorithm can get a fault diagnosis accuracy of86.67%. This suggests thatwhen we use SWFCM algorithm divide transformer fault samples, the sample’s numbershould be large and uniform as far as possible.