Generator Stator Partial Discharge Pattern Recognition Method Based on BOVF and Active Learning QBC-Catboost

Partial discharge(PD) of generator stator bars is an important basis for generator insulation diagnosis. The traditional supervised classification model requires a large number of labeled training samples, which has high requirements on sample quality, and it is difficult to construct and optimize the performance of conditional classifiers with few labeled samples. To solve the problems, this paper proposes a feature extraction method based on bag of visual features(BOVF) and active learning QBC-Catboost partial discharge pattern recognition model. First, the model constructs BOVF of PRPD through bag of visual words model, and then uses active learning query by committee (QBC) strategy and Catboost classifier to jointly measure sample value and complete value retrieval. Finally, BOVF -QBC-Catboost model realizes labeling cost constraint and recognition performance optimization. Based on laboratory conditions, five typical defective of stator bars were designed and defect samples were obtained to verify the validity of the model. The accuracy of BOVF can be improved by up to 5.3 % compared with the statistical characteristics of discharge. The recognition accuracy of QBC-Catboost model with 25% sample labeling rate can reach 98.4%, and the difference of recognition accuracy between QBC-Catboost model and fully labeled random sampling model is only 1.0%. The experimental results show that BOVF and QBC-Catboost recognition model can effectively mine discharge characteristics, complete value sample retrieval, effectively constrain the training cost, and realize the construction of high-performance classification model.