Research On Pattern Recognition Statistical Parameter Methods Of Partial Discharge Models In Oil-filled Insulation

Partial discharge (PD) in the oil-filled insulation equipment is considered as a major cause of insulation degradation. Internal insulation partial defect and the relevant discharge development degree could be find out in time by on-line monitoring PD signal and recognizing PD type, so the coming faults is prevented. According to the difficult problem of PD pattern recognition, this paper analyzed the statistic method and artificial neural network method, simulated the typical PD artificial defects and obtained lots of PD sample data, and different methods were compared.The PD parameter extracting methods of the fingerprint charts were researched, and a new method was put forward that used Weibull model to analyze PD pulse amplitude distribution spectrum. Weibull parameter scale parameter and shape parameter were used as new character. After comparing the parameter estimate methods, this paper suggested that use the least square methods to calculate the double-parameter Weibull distribution and use the optimized model to calculate the mixed Weibull distribution. Particular analysis was done according to the single PD resource and mixed PD resources, and Weibull parameters' distributions corresponding PD types were summarized. PD pulse phase distribution spectrum was compared with spectrum, and 95% fiducial intervals of statistic operators were also summed up. The clustering analysis results have proved that Weibull parameters could represent the PD types and the degree of PD development very well. When applying the artificial neural network to recognize the PD patterns, merits and defects of BPNN was researched, the structure ,learning arithmetic and training arithmetic were improved, and recursion NN with windage cell was designed. Aim at the problem of NN input choice, a new way which use ,,Sk and Ku to generate input vector was brought forward. From the results of comparison with data array input method, it showed that the former has better effect and simplified the NN structure and shortened the learning and training time.