Feature Selection And Deep Learning Based Partial Discharge Pattern Recognition Of High-Voltage Cables

High-voltage cables are an important part of the city’s power supply network.Their safe and stable operation is closely related to the normal production and life of the people.Because of the long-term high-voltage and high-current operation,high-voltage cables are prone to operational failures due to external force damage,equipment defects,and invasion of water trees.Partial discharge(PD)monitoring can collect PD signals,analyze and process signals,which is useful to the timely detection of cable faults.PD pattern recognition is one of the important aspects of PD monitoring.The PD recognition is difficult.On the one hand,the PD signal is mixed with the background noise.On the other hand,the similarity between different PD types is very high,which increases the difficulty of the PD identification.The study on PD recognition is helpful to timely and accurately determining the type of fault.It is of great significance to ensure the safe and stable operation of the cables.On the basis of summarizing the current research status of PD pattern recognition,this paper clarifies the problems and challenges faced in the field of PD recognition.Based on the existing research deficiencies,the data acquisition,data augmentation,feature selection,and pattern recognition are studied.The main findings are as follows:(1)Aiming at the problem of long time and low efficiency of mass PD processing,an automatic preprocessing system for PD data is proposed and a corresponding system is developed.The number of original samples is huge.It is a big challenge to efficiently extract and process PD and interference signals from massive samples.The automatic preprocessing system in this paper first divides the signals into different types through multi-dimensional feature classification,and then uses the phase resolved PD pattern recognition method based on the K-Means clustering algorithm to automatically determine whether the signal belongs to PD or interference.After the judgment,the signals are added to the sample database.When a new signal arrives,it first looks for a similar signal by the sample database to determine the category,and if there is no similar signal,it continues to be identified by the phase resolved PD pattern recognition method.The system gives confidence at the judgment stage,and the signal with low confidence level needs to be judged by the expert assistance analysis system before it can be added to the database.The automatic preprocessing system on the one hand automatically processes PD data and statistically determines whether the signal is a partial discharge.On the other hand it constructs a historical sample database of PD and interference signals,which is useful to facilitate follow-up PD research work.(2)For the problem that the training data is not sufficient,a data augmentation method based on random noise and generative adversarial network is proposed.Some complex pattern recognition methods(such as deep neural networks)have many parameters and require large training data.If the amount of data is insufficient,the model is easy to be overfitting,so this paper studies the data augmentation method for PD.In this paper,data augmentation methods based on random noise and generative adversarial network are adopted to generate new PD data.This paper uses pattern recognition methods such as support vector machine and visualization methods to verify the effectiveness of data augmentation to improve the accuracy of PD recognition.The random noise method lifts 0.99% average accuracy using support vector machine,and enhance 0.96% average accuracy using random forest.The generative adversarial network lifts 1.71% accuracy using logistic regression and 0.53% accuracy using support vector machine.The results fully demonstrate the effectiveness of data augmentation methods based on random noise and generative adversarial network.(3)Aiming at the problem of too many feature parameters which makes it difficult to determine the effective features,a feature selection method based on random forest is proposed.There are many characteristics of PD,and the dimension disaster caused by the high feature dimensions reduces the training efficiency and affects the training accuracy of the model.Therefore,this paper studies the feature selection method for PD.This paper proposes a feature selection method based on random forest to construct new features and find effective new features.It constructs 1235 PD features and ranks the PD features to find effective features.In order to enhance the credibility of feature selection results,this paper uses pattern recognition methods such as support vector machine and visualization method to verify the results of feature ranking.The verification results show the effectiveness of PD feature selection based on random forest method.(4)Aiming at the problem of low recognition accuracy due to the high degree of similarity of different PD types,a PD identification method based on convolutional neural network is proposed.This paper uses convolutional neural network to identify 5 types of total 3500 PD signals,and further studies the influence of network structure,activation function,optimization algorithm on the accuracy of PD pattern recognition.The results show that the convolutional neural network effectively improves the accuracy of PD recognition.Compared with the BP neural network and the support vector machine,the accuracy increases by 4.29% and 3.01% respectively.It can better distinguish the type 2 and type 3 which are difficult to identify.