Research On Method Of Partial Discharge Fault Feature Extraction

Aiming at the problems of non-stationarity and non-linearity of transformer partial discharge signals,as well as insufficient deep learning and strong subjectivity when traditional machine learning extracts signal features,this thesis proposes a Symmetrized Dot Pattern image feature fusion-volume Partial discharge feature extraction method based on Convolutional Neural Network.In the process of feature extraction,this method mainly uses neural network as a recognition tool,and uses the fusion features of signals as the input basis of the network in the process of state recognition.Finally,a transformer partial discharge state recognition model with fusion feature learning is established.This model can fully extract the effective features in the partial discharge signal in an adaptive manner,and better realize the functions of feature extraction and state recognition.First,the resonance sparse decomposition is used to extract the high resonance and low resonance modal components of the partial discharge signal.The decomposition is based on the nature of the partial discharge si gnal.By changing and optimizing the parameters of the resonance factor,the deficiency of signal feature overlap in signal processing is improved.This method effectively eliminates the non-stationary characteristics,impact characteristics and random osc illations of the signal.All aspects of the signal are extracted more effectively.Secondly,on the basis of resonance sparse decomposition,SDP image processing of high and low resonance components is carried out.This thesis proposes a method of partial discharge feature extraction based on Iradon image transformation.Both are about SDP and Iradon image.The characteristics of the signal are expressed more vividly.The feature information fusion SDP image can display the characteristics of the output signal more clearly,intuitively and comprehensively.The Iradon transform is also the form of feature images,thus greatly improves the distinguishability between features in different states.Finally,this thesis proposes a method for SDP image and Iradon i mage recognition based on convolutional neural network.Using the powerful calculation and recognition functions of CNN,learn different state features from actual partial discharge signals.The method can reduce subjective discrimination and errors in the learning process for achieving feature extraction and state recognition.