Recognition Of Transformer Discharge Phase Pattern Based On Multiple Neural Network Methods

Partial discharge of transformer is the main cause of transformer insulation failure.Effective identification of different types of partial discharges is conducive to fault location and elimination.Partial discharge phase distribution analysis(PRPD)is currently the most widely used and effective partial discharge pattern recognition method.This method is based on PRPD atlas for feature extraction and discharge type classification,but the commonly used feature extraction methods are based on detection of raw data.Due to the different data formats used by the testing equipment and the confidentiality of the data among different companies,the incompatibility of partial discharge diagnostic methods has resulted.The PRPD atlas has relevant international standards,and the rendering method is uniform.The pixelation of the atlas is directly identified,and the feature extraction of the original data is avoided,which is beneficial to the mobility of the identification method.At present,the direct identification of PRPD maps is limited to experienced experts,and the workload is large.In order to solve this problem,this paper uses image recognition algorithms to perform machine recognition on various types of PRPD maps,which not only improves efficiency,but also has a strong generality of the final model and higher recognition accuracy.The following research has been done specifically:(1)A convolutional neural network based on multi-layer feature fusion is designed and used for PRPD q-φ atlas recognition.On the basis of the original network structure,the outputs of all pooling layers are all input to the fully connected layer,which preserves the original features of the input map to the greatest extent.Secondly,the traditional CNN’s pooling strategy is improved.The largest two mean pooling is used,and the mean of the two largest pixels in the pooling window is used as the lower layer input to further improve the accuracy of the discharge type recognition.(2)An improved residual network(ROR-3 Res Net)based on a mixture of activation function upgrades and ROR cross-layer connections was designed and used for PRPD discharge fingerprint identification.The SELU unit was first introduced to change the status quo of the death of negatively gradient neurons due to the negative semi-axis constant of the traditional activation function.Second,based on the basic residual network sequential stacking structure,a multi-level shortcut connection was introduced to make T he "bad" fit is changed to the "residual residual" fit,the training speed is faster,and the recognition accuracy is higher.(3)The pre-processing methods of the PRPD q-φ spectrum and fingerprint spectrum are studied.The two network models are trained b y using the pre-processed samples.Finally,the validity of the two network models is verified using the actual fault field detection data.Experiments show that the improved algorithms proposed in this paper for basic CNN and Res Net perform well in partia l discharge pattern recognition based on the PRPD method,and have great application prospect in other image recognition scenarios.