Study On Partial Discharge Pattern Recognition In Transformers Based On Grayscale Image And Affinity Propagation Algorithm

During the operation of the transformer,its internal insulation fault is one of the main causes of the accident.Partial discharge(PD)in power transformers is closely related to the insulation defects.Different from PD fault diagnosis based on threshold,pattern recognition judges the character and types of PD through data acquisition,feature extraction and classification decision,and then diagnoses the insulation defects of transformers based on the character and types of PD.PD pattern recognition is not only the basis for evaluating the insulation condition,but also an important means for determining insulation defects,which plays a key role in monitoring the operating state of transformers.In this paper,three internal defect models of power transformers are designed to simulate surface discharge,corona discharge and suspension discharges.Ultra-high frequency(UHF)and Acoustic Emission(AE)are used to measure PD signals.The discharge experiment of a single model and the mixed discharge experiment of three models are carried out respectively.The distribution characteristics of all PD signals are analyzed based on phase resolved partial discharge(PRPD)patterns.The distribution of UHF signals of surface discharge,corona discharge,suspension discharge and mixed discharge has obvious differences and features.19-dimensional feature parameters belonging to moment feature,fractal feature and textural feature are extracted from the grayscale image of PRPD patterns and used for PD pattern recognition.The effects of the number(L)of gray scales on textural features and the selection of non-scale range of fractal feature are analyzed respectively.The feature parameters are appraised according to the probability distribution of the feature parameters.Moment feature is better than textural feature and fractal feature at distinguishing PD types,and textural feature has the worst distinction.yo+ is the best feature parameter of UHF grayscale image,and xo+ is the best feature parameter of AE grayscale image.In order to reduce the burden of high-dimensional data on classifier training,principal component analysis(PCA)is used to reduce the dimensions of 19-dimensional feature parameters.The new feature parameters obtained by PCA are the most difficult to distinguish the mixed data of surface discharge and corona discharge.The influences of the preference p and the damping factor ? on clustering of the affinity propagation(AP)algorithm are analyzed based on Iris dataset.Kernel affinity propagation based on particle swarm optimization(PSO-KAP)is proposed for the problem that AP algorithm cannot cluster the data with specific structure and is only suitable for the data with compact hyperspherical structure.PSO-KAP classifier is designed for PD pattern recognition.Cross-validation is used to calculate recognition rates of PSO-KAP,AP,k-means,back propagation neural network(BPNN)and least squares support vector machine(LSSVM)for Iris dataset and four types of PD.The performance of PSO-KAP is better than that of AP,k-means and BPNN.which is close to the performance of LSSVM.But the misrecognition rate of LSSVM for explicit data is higher than that of clustering algorithm.The recognition rates of PD by PSO-KAP is more than 98%.Studies have shown that PSO-KAP is better than AP and k-means for the scattered data.Compared with the LSSVM classifier,PSO-KAP doesn't show its advantage based on PD data.The UHF signal is more advantageous than AE signals for the identification of PD typesThe feature extraction of grayscale image of PRPD patterns is studied.The results show that moment feature.fractal feature and textural feature of grayscale image have excellent application value for PD pattern recognition.A new solution for PD pattern recognition is proposed.This method achieves ideal recognition results based on experimental data,and it has a certain application prospect.