Research On Intelligent Detection Method Of Partial Discharge Fault For Insulated Overhead Conductors

Insulated overhead conductors has basically replaced the traditional steel-cored aluminum stranded wire,with the external insulation layer to protect the people's life and industrial production of electricity demand.However,the existence of the insulation layer makes the traditional fault detection device can not be used normally,which brings a huge obstacle to the fault detection of insulated overhead conductors.Because the failure of insulated overhead conductors is often accompanied by Partial Discharge(PD),the fault of insulated overhead conductor is often judged by the partial discharge of insulated overhead conductor.Aiming at the problem of fault detection accuracy of insulated overhead conductors,this paper analyzes and processes the data by means of signal decomposition and noise filtering,an improved feature screening method and an improved deep learning method are proposed,and the proposed method is verified and analyzed by experiments.The work of this paper is mainly summarized in the following three aspects:(1)A noise filtering method for pulse voltage signal based on time series decomposition is designed.For the background noise in the pulse voltage signal data of insulated overhead conductors,a noise filtering method combining with time series decomposition is designed in this paper.The characteristics of voltage signal peak and noise peak are found according to the variation of peak gradient in residual components.By analyzing the extracted peak information,five kinds of peak characteristics are calculated.Among them,the RMSE with sawtooth wave of peak is the most important characteristic,which contains the special properties of partial discharge phenomenon at the peak,which greatly improves the characterization ability of signal characteristics.(2)An ATPO-LightGBM feature selection algorithm was proposed.Feature screening can improve the upper limit of model performance and make the model more accurate.In the process of probabilistic results binarization of the model output,the optimal probability threshold value is calculated by the probability estimation value and the real classification of the training set,which is more robust than the original transformation method with the threshold value of 0.5.It is found that the RMSE with sawtooth wave of peak is a very important feature,and its feature importance is much higher than other features.By comparing the model precision before and after feature screening,it is proved that feature screening can improve the intelligent recognition method.(3)A partial discharge detection method with deep learning using improved LSTM is proposed.A Bi-LSTM network combining attention mechanism(ATT-BLSTM)is proposed based on the strong learning ability of RNN to sequence data.The attention mechanism is added between the encoder layer and decoder layer of Bi-LSTM to give different weights to the hidden features.The importance of hidden features is evaluated to improve the efficiency and performance of the model.By contrast experiment,the reliability of the deep learning method is proved,and it is proved that both feature screening and attention mechanism play key role in feature engineering.