Study On Pattern Recognition Of Partial Discharge Signals In Electrical Equipment Based On Convolutional Neural Network

The insulation performance of power equipment directly affects the safe operation of the whole power grid.Partial discharge(PD)is the main sign and manifestation of insulation deterioration of power equipment.The identification of PD type is conducive to the evaluation of insulation status of power equipment,and helps the operation and maintenance personnel to find the insulation defects of power equipment in time.In this paper,the partial discharge signal sample data generated by artificial insulation defect model is used to study the pattern recognition of partial discharge signal of power equipment insulation.The main contents of this work are as follows.A time-frequency analysis method of PD signals based on variational mode decomposition(VMD)and Choi-Williams distribution(CWD)is proposed.Firstly,a series of band-limited intrinsic mode functions(BLIMFs)are obtained by VMD of the PD signal superimposed by multi-components.Then,CWD of each BLIMF is calculated.Finally,the complete VMD-CWD analysis results of PD signal are reconstructed by linear superposition of the CWD results of each BLIMF.The experimental results show that VMD-CWD time-frequency analysis method can not only effectively suppress the cross interference,but also ensure that the spectrum has good time-frequency aggregation,and can clearly reflect the change process of PD signal in time-frequency domain,which is conducive to the subsequent feature extraction and pattern recognition.A feature extraction method of PD signals based on cross-layer feature fusion and optimization convolutional neural network(CFFO-CNN)is proposed.On the basis of the obtained VMD-CWD time-frequency spectrum,the convolutional neural network is used to realize the automatic feature extraction.Firstly,VMD-CWD is output to image format,and then graying processing and image compression are carried out.Secondly,CNN is constructed and the network structure and initial parameters are determined.Then,CFFO-CNN is obtained by introducing cross-layer connection to fuse deep and shallow layer feature information,and the network is improved by optimization algorithm.Finally,the autonomous feature extraction and recognition of VMD-CWD images are implemented by using CFFO-CNN.The results prove that this method can effectively extract the deep and shallow feature information of the VMD-CWD images of PD,avoid the complexity and subjectivity of the manual feature extraction process,and the obtained feature quantity has a better recognition effect.A pattern recognition method based on Hilbert marginal spectrum image and deep residual network is proposed.Firstly,the VMD algorithm is used to decompose the PD signal,and the corresponding Hilbert marginal spectrum of each BLIMF is constructed by Hilbert transform.Secondly,save the marginal spectrum as an image,and the image is preprocessed as the input data.Finally,the deep residual network is used to automatically learn the intrinsic characteristics of image pixel data to complete the pattern recognition of PD signals.The experimental results show that the recognition method used in this paper has a high correct recognition rate.