Study On Methods Of Recognition Feature Extraction For Partial Discharge Images

Partial discharge (PD) inside insulation is considered as one major cause of insulation degradation in electrical equipment and attached importance to the safety and reliability of running electrical equipment. Auto-recognition to discharge types in on-line PD monitoring system could be used to find out internal partial defects and the relevant discharge development degree in time,and then prevents equipment from the coming faults. According to the requirements to PD pattern auto-recognition,this paper studies systematically the basic theories and realizable methods for auto-recognition ofPD gray intensity image:(1) To meet the requirements of on-line PD monitoring for transformer,several discharge models are designed and the relevant experiment methods are projected. With discharge model tests,a lot of discharge sample data is acquired. We established the foundation for studying on PD gray intensity image and BPNN classification.(2) Based on PD fractal features,the modified differential box-counting (MDBC) method was studied to extract fractal features of PD. In addition,three kinds of fractal feature characterization and extraction methods,including the 2nd generalized dimensions of original PD images and fractal dimensions of high gray intensity PD images are studied. By calculating fractal feature of specific discharge sample,certain differences between fractal features of different discharge are demonstrated.(3) The statistic character of PD gray degree images are studied. Firstly,the square character of PD gray degree images is studied,which is used in describing geometry character of PD gray degree images;secondly,the relative statistic character parameters of PD gray degree images is studied. By calculating statistic character parameters of discharge samples,we can also find certain differences between fractal features ofdifferent discharge exist.(4) Te classifier with back-propagation neural network (BPNN) is designed. The comparatively high recognition correctness probability is achieved in classification to original PD images. The above research results show that the proposed PD recognition features set together with BPNN classifier,can be effectively used in PD pattern recognition. Satisfying recognition results are achieved by using the above mentioned method..