Research On Partial Discharge Pattern Recognition Of Typical Defects Of GIS Basin Insulator Based On Ultrasonic Method

GIS is widely used in substation of various voltage levels due to numerous advantages,such as small footprint and high stability.Basin insulator is the core insulation component of GIS.Affected by many factors in the process of manufacturing,installation and operation,there may be a variety of insulation defects on its surface and inside.Therefore,partial discharge is prone to occur in operation,which accelerates the insulation damage process of basin insulator.There are few studies on the identification of partial discharge of basin insulator,so it is necessary to carry out relevant studies.In this paper,the partial discharge experiment platform of typical defects of basin insulator is established,and a large number of partial discharge ultrasonic signals are collected.The feature parameters and feature sequences of the signals are extracted based on the local mean decomposition(LMD),and the classifier is used for training and testing to verify the feasibility of the proposed classification and recognition method.The main work and conclusions are as follows:(1)Based on the actual working conditions of GIS basin insulators,a partial discharge model with typical defects of basin insulators is designed and fabricated in the laboratory.By controlling the epoxy resin curing condition,small basin insulator experimental samples with crack,bubble,insert burr and surface metal particles defects are fabricated.An ultrasonic partial discharge detection experimental platform is built,and experiments are carried out to collect a large number of ultrasonic signals of partial discharge of different defects.(2)The results of decomposing the partial discharge simulation signal using standard LMD show that the standard LMD suffers from modal mixing and generates excess spurious components when decomposing the partial discharge simulation signal.The standard LMD is optimized by means of positive and negative paired Gaussian white noise assisted decomposition and introduction of improved LMD algorithm.Its performance improvement is verified by decomposing the partial discharge simulation signal again,which can be used for partial discharge signal decomposition.(3)A feature parameters extracting method of partial discharge signals is proposed and then classification is achieved using support vector machines(SVM).The optimized and improved LMD is used to decompose the partial discharge signal and obtain the components.The energy percentage,energy entropy and Renyi entropy of the main components are used as the feature parameters and brought into the SVM for classification and recognition,and the correct rate reaches 96%.Varying the train set sample size for classification,the results show that the proposed method can still achieve 94.5% correct rate even with small samples,and the correct rate gradually increases when the train set sample size increases,but the total time consumed increases greatly,which proves the effectiveness of the extracted feature parameters and the advantage of support vector machines for small sample classification problems.(4)A extraction method of partial discharge signal feature sequences is proposed and then classification is achieved using long short-term memory neural network(LSTM).Each main component is divided into equal length signal segments,then the energy share,Renyi entropy,and Hurst exponent of each signal segment are extracted,and the corresponding feature sequences are synthesized according to the temporal order and brought into LSTM for classification and recognition,and the correct rate reaches98.25%.The results of varying the sample size of the train set for classification show that the correct rate of the proposed method exceeds 95% for medium or larger samples,and the time consumption is significantly reduced compared with SVM,while the recognition effect is affected when the sample size is small,which verifies the effectiveness of the extracted feature sequences and reflects the efficiency of LSTM in the recognition process of large sample training and the limitation in dealing with small sample problems.