Research On GIS UHF Partial Discharge Pattern Recognition Based On Deep Learning And Multi-information Fusion

Gas insulated switch-gear has been widely used in the field of power transmission and distribution by virtue of its good insulation performance,high power reliability and small footprint.With the increase number of GIS put into operation in the power system,failures caused by insulation defects during equipment manufacturing,transportation,assembly and long-term operation are also gradually increasing.Partial discharges caused by different types of insulation defects have different damage to equipment insulation and the risk of breakdown.The identification of defect types has important guiding significance for handling defects and making maintenance decisions.Due to the diversity and complexity of partial discharge phenomena,the current partial discharge samples established through typical laboratory models and defect classified by single pattern recognition algorithm are not effective in engineering field applications.To this end,this paper carried out experiments to expands the partial discharge sample library,integrated the BP neural network model and the convolutional neural network model,and studied GIS UHF partial discharge type recognition method based on complementary weighted fusion of multiple models.Committed to improving the accuracy of partial discharge pattern recognition.In this paper,five typical insulation defect models were designed.While retaining their unique typical structural features,partial discharge experiments were carried out by changing the size and shape of defects and applying different voltage levels.A large amount of partial discharge sample data was obtained by using UHF detection method,and a relatively complete partial discharge sample library was constructed.In order to use the information of discharge data more comprehensively,three images of PRPD,△t,and Δu reflecting the phase statistical distribution of discharge pulse,the time interval between adjacent pulses and the information of the applied voltage difference when adjacent discharge pulse occurs were constructed,a convolutional neural network model was built to classify and recognize three types of images;at the same time,26 kinds of characteristic parameters including numerical statistical characteristic parameters,fractal characteristic parameters and image characteristics of UHF partial discharge sequence signals were extracted in parallel and trained BP neural network.On the other hand,in order to overcome the shortcomings of a single classifier,a GIS partial discharge type recognition method combining two algorithm models of convolutional neural network and BP neural network was proposed,which combines the image information and statistical parameters constructed by partial discharge pulse sequence,Finally,when identifying the type of discharge,the recognition results output by the four models of the sample to be tested were weighted and fused,and the final recognition result of the discharge type was obtained.The research results show that the discharge type recognition algorithm based on deep learning and multi-information fusion proposed in this paper,through mutual corroboration of various models can make full use of the multi-dimensional information of partial discharge signals,overcome the problem of low recognition rate of certain defects by a single diagnosis method,and effectively improve the accuracy and engineering practicability of partial discharge type recognition.