Intelligent Partial Discharge Diagnosis Based On Acoustic-Electric Fusion

With the development of the power grid,the stable and reliable operation of power equipment has higher requirements,and it has also driven the gradual maturity of partial discharge charging detection technology,such as UHF detection method and ultrasonic detection method,which have been obtained at the substation site.Wide range of applications,which also promotes the continuous development of state monitoring of transformers under operating conditions.At present,domestic and foreign scholars have proposed various pattern recognition algorithms based on different detection methods for pattern recognition of partial discharge of transformers.However,there is little comprehensive analysis of the data obtained from a large number of different sources,and the utilization of data is not high.Therefore,this paper proposes an acoustic-electrical fusion transformer partial discharge intelligent diagnosis method,which can combine the partial discharge UHF and ultrasonic signals of the transformer to make a diagnosis,which not only can more effectively utilize multi-source data,but also improve the recognition accuracy.The experimental data of the paper was collected on the true transformer of the laboratory and detected by UHF sensor and ultrasonic sensor respectively.Each detection method collects more than 2,600 pieces of data on four kinds of defect models: tip corona,suspension discharge,insulating air gap and creeping discharge,and then realizes pattern recognition through machine learning,deep learning and acoustic-electric fusion intelligent algorithm.A high recognition accuracy rate.A three-dimensional phase-resolved pulse sequence(PRPS)map and its corresponding two-dimensional map are generated for the detected UHF partial discharge signal,including maximum discharge amount-phase,discharge number-phase,and average discharge amount-phase distribution..The two-dimensional distribution of statistical parameters,including skewness and kurtosis,are obtained to obtain 21-dimensional eigenvectors.This paper proposes an online sequence extreme learning machine(OS-ELM),which implements the incremental learning improvement algorithm of the traditional extreme learning machine(ELM),which not only improves the recognition accuracy,but also has significant significance in training time and generalization performance.improvement of.Compared with traditional ELM,support vector machine(SVM)and BP neural network,the recognition accuracy rate is increased by 2.5,4.9 and 23.4 percentage points respectively;only 1/10000 of SVM and 1/5000 of BPNN are needed in training time.For the partial discharge ultrasonic signal,because its feature extraction is more complicated than UHF signal,and because of the advantages of deep learning network in learning feature information,this paper proposes to use deep learning network,including convolutional neural network(CNN),recurrent neural network.(RNN)and deep neural network(DNN)to learn the characteristic information of the ultrasonic signal and realize pattern recognition.Firstly,the ultrasonic signal is downsampled,and the Mel frequency cepstral coefficient(MFCC)is combined with the characteristic parameters such as chromaticity diagram and spectral contrast map to train CNN,RNN and DNN respectively,and the recognition results and performance are compared.It is proved that CNN is The advantages of ultrasonic recognition finally achieved an accuracy of 92.9% and an AUC score of 99.82%.Considering that the UHF signal or ultrasonic signal is used to deal with multi-source data too lowly,this paper proposes an intelligent diagnostic system for acoustic-electric signal fusion.Based on the fusion characteristics of multi-source signals,the machine learning and deep learning networks are trained respectively to obtain the recognition results.Then based on Choquet fuzzy integral,the classification results of each classifier are fused to obtain the final classification result,which greatly improves the recognition accuracy rate,especially in the sample type with poor recognition effect of single classifier.The lifting effect is very obvious.For the SVM and CNN identification of the correct air gap discharge type of only 75% and 87%,the acoustic and electrical fusion intelligent diagnosis system finally achieved a correct rate of 90.89%.The correct recognition rates for suspension and surface discharge are as high as 95.27% and 96.31%,respectively.