Study On Denoising And Feature Extraction Approaches For Partial Discharge Signals In Electrical Equipment

The insulation of electrical equipment is one of important factors related to the safety and stable operation of an power system.Partial discharge?PD?is the important symptom and manifestation of insulation deterioration in electrical equipment.It is also a sign of deterioration of insulation integrity of electrical equipment.Due to the different reasons for the internal insulation degradation,the type of PD exhibited by the device is also different.Therefore,the recognition of the PD types is of great significance for further study the internal insulation and the evaluation of electrical equipment status.In addition,the strong on-site noise suppression and the efficient feature extraction of detection signals for pattern recognition are key parts for the analysis of PD signals.Based on the analysis of the characteristics of PD signals,t he denoising method,the feature extraction method and pattern method of PD signals are studied.The main work of this paper is as follows:A wavelet package denoising method based on complete ensemble empirical mode decomposition?CEEMD?and permutation entropy?PE?is used to denoise the partial discharge signals.Based on the CEEMD of noisy signals,the decomposed modal components are arranged according to the values of PE,and the modal components that need to be discarded or further decomposed are determined.For those components that need further processing wavelet package transformation is performed,and the decomposed component signals are reconstructed to obtain the PD signals after denoising.Simulation and experimental results show that the propos ed method achieved an ideal denoising effect,which verified the effectiveness of the proposed method and facilitated the further processing of the pattern recognition of PD signals.A feature extraction method based on 2D variational mode decomposition?V MD?and Hilbert transform is proposed in this paper.Firstly,the PD grayscale images are formed by the measured PD samples.Then,the PD grayscale images are decomposed by 2D VMD algorithm to obtain the modes on different center frequencies.Finally,quaternion Hilbert transform are used to obtain the corresponding characteristic graphs,and the corresponding eigenvector of each PD sample is come into possession by the texture features of these graphs.The experimental results show that the extraction method has a high correct recognition rate,and it verifies the feasibility of the feature extraction method.In addition,the 2D VMD-Hilbert method also provides a new idea for the spectrum analysis on PD signals.A method based on Hilbert marginal spectrum and SAE-DNN is proposed in this paper to recognize PD types.Firstly,PD signals are dealt with variational mode decomposition?VMD?.And these obtained modes are used to construct corresponding Hilbert marginal spectrum.Secondly,a Hilbert marginal spectr um is taken as the an input vector.And SAE can learn the inherent features and extract the succinct expressions from original data automatically.Then,the results obtained by SAE are used to initialize DNN.And DNN is trained by a large numbers of sample s.Simultaneously,in order to speed up the convergence in the processes of learning of SAE and DNN,the network is optimized with the adaptive-step learning rate and updated with the weight parameters.Finally,DNN is trained to identify the PD types of samples.The experimental results show that results based on SAE-DNN can achieve a higher accuracy and less time.A method to identify unknown PD types based on improved support vector data description?SVDD?algorithm and Mahalanobis distance is presented in this paper.And a dual threshold method based on Otsu criterion is proposed to determine the types of PD samples.Firstly,PD samples were collected from artificial defects models and extracted feature vectors to constitute sample sets.Secondly,the SV DD algorithm was used to solve the center and the radius of the hypersphere of the training PD samples.Then,the double threshold R₁ and R₂ were set according to the Otsu criterion,and the feature space was divided into different regions.Finally,according to the classical criterion and Mahalanobis distance,the types of PD of the test samples were determined.The experimental results show that the accuracy of recognition obtained by the method proposed in this paper is high,which verifies the feasibility of the method.