The Method And Experiment Research Of PD Pattern Recognition Based On Fuzzy Clustering Theory

Partial discharge (PD) is considered as one major parameter which assesses the insulation degradation of high voltage electrical apparatus. The insulation state directly determines the stability of high voltage electrical apparatus. If only the operating states of high voltage electrical apparatus on-line are monitored and PD signals are sampled, the types of discharge can be identified. The insulation weakness and the discharge developing degree of the high voltage electrical apparatus can be discovered in time, which can stop accidences. Therefore, a study on pattern recognition of PD has both important academic meaning and engineering value.Based on a great deal of research literatures which have been published in domestic and aboard on pattern recognition, many simulating PD experiments are made. This thesis studies the PD pattern recognition techniques which based on fuzzy clustering, extracts the feature of the PD pulse signals and comprehensive analysis compares the fuzzy clustering methods. It provides effective methods for PD pattern recognition technique. The main works in this paper are as follows:Three kinds of PD models are made according to the mechanism of generating PD, which include point-point electrode system, point-plate electrode system and sphere-plate electrode system. Making use of the method which bases on fuzzy inference system to wipe off the colored noise after sampling PD pulse signals by DSO2902 data acquisition device. Carrying through pattern recognition to PD signals by transitive closure method of fuzzy equivalent relation and fuzzy c-means clustering algorithm when the moment feature of PD image gray and image fractal feature are extracted, and the differences of their clustering results are analyzed from algorithm's point. Although clustering effect of fuzzy c-means algorithm is better than transitive closure method, the principal defects of fuzzy c-means algorithm has been convergence speed slow and not cut for large discrepancy of every class sample number. Thereby based on fuzzy c-means algorithm having limitation of equal demarcation trend for sample sets, distributing density size of sample dot is regard as weighted value, a weighted fuzzy c-means algorithm is proposed and PD signals pattern recognition rate is above 90%. This algorithm tests the type IRIS data sets of international recognized compare clustering result performance. The result analysis indicates that algorithms not only have certain extent overcome limitation of fuzzy c-means algorithm, but also have fine central position than that of the fuzzy c-means algorithms, furthermore, favorable convergence.