Research On Incipient Cable Failures Identification Technology Based On Residual Network

During the long-term operation of the cable,the aging of the line is easy to form local insulation defects,which will lead to various types of faults.The incipient cable fault is an intermittent arc fault caused by defects in local insulation of cable,which will accelerate the deterioration of cable insulation and lead to permanent damage of cable in serious cases.Due to the short duration and self-recovery of the incipient cable fault,relay protection devices often regard it as a transient disturbance and ignore its potential harm.At the same time,there are other forms of transient disturbances in power grid operation,which makes the identification of the incipient cable failures more difficult.Based on the above problems,this paper proposed an effective fault classification method to realize the accurate identification of the incipient cable fault.The main research work can be summarized as follows:(1)the relevant characteristics of the incipient cable fault were studied and analyzed.According to the time-varying and nonlinear characteristics of the fault resistance of the incipient cable fault,the arc model was used for equivalent analysis.Based on the improved cybernetic arc mathematical model,the incipient cable fault model was built in PSCAD.The transient waveform characteristics of multi-cycle early fault and half-cycle early fault were compared and analyzed in two grounding systems respectively,which verifies the effectiveness of the arc model.Through the control parameter comparison experiment,the parameter modification basis was provided for the generation of the incipient cable fault data;In order to increase the diversity of fault samples,three fault interference models were introduced for subsequent fault classification research.(2)In order to extract fault features more accurately and improve the performance of fault classification model,the fault feature extraction method was studied and verified.The simulation model of 10 k V distribution network was built in PSCAD,and multiple types of fault data were generated through batch simulation of automation library.Aiming at the defects of large dimension of original fault sample data and easy loss of mutation features,a feature sample construction method based on stationary wavelet transform was proposed.The sample feature matrix was constructed by extracting the feature quantity of original samples,so as to realize more accurate feature mining of fault samples;In order to verify the effectiveness and superiority of the proposed method,three machine learning algorithm models were built,and the classification effects of the original samples and feature samples were compared.It was proved that the proposed method could greatly improve the classification performance of the network model.(3)Aiming at the limitation that machine learning algorithm could not mine the deep-seated features of faults,a fault identification method based on deep learning algorithm was proposed.The fault classification models based on the convolution neural network and the residual network were constructed respectively.On the basis of the residual network,in order to fully mine the timedomain and frequency-domain information of faults,the residual block was improved based on the idea of separation convolution,and an improved residual network fault classification model was built.The evaluation results showed that the improved residual network model could achieve faster convergence and higher classification accuracy.At the same time,the anti-interference analysis results also proved that the improved residual network model had stronger robustness.