Recognition Method of Partial Discharge Spectrum Feature in Generator Stator Bar Based on Random Forest

On-line monitoring and fault diagnosis of partial discharge (PD) in generator stator bar is of great significance for fault warning, fault locating, and helping unit maintenance. There are problems that some different types of partial discharges have high similarity and the calculation time of pattern recognition method is excessively long, which needs a high-precision and fast recognition method. Therefore, a partial discharge recognition method of stator bar based on random forest is proposed. Six types of defective stator bars are fabricated, and ultra-high frequency antenna is used to obtain partial discharge signals. The phase resolved partial discharge (PRPD) spectrum is compared to extract a feature called amplitude asymmetry. Moreover, the defect type is identified by the random forest method, and the feature importance is calculated to select effective features. Finally, from the viewpoint of visual feature similarity, the random forest is compared with the traditional back propagation (BP) neural network to verify the effectiveness. The results show that the random forest algorithm can effectively identify the partial discharge from artificial defects in stator bars with an overall correct identification rate of 93.33%. After halving the number of features, the accuracy of random forest is 10.83% higher than that of BP neural network. Feature selection has a negligible effect on the accuracy of random forest, whereas the recognition efficiency is greatly improved. The recognition accuracy and calculation time of random forest are significantly better than those of neural network in a small number of features. The validity of the amplitude asymmetry is ranked in the first third of all features, which can be promoted further.