Deep convolutional neural network for partial discharge monitoring system. (June 2023)

• Record Type: Journal Article
• Authors: Srivastava, Rajat; Avasthi, Vinay; R․, Krishna Priya
• Is Part Of: Advances in engineering software. Volume 180(2023)
• Publication Date: 2023-06
• Subjects: Electrical cable -- Partial discharge condition monitoring -- Proposed PCA- diminished technical indicators -- Feature Extraction -- Optimization; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103407
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: Proposed a novel data-driven approach to recognize the condition PD pulses of power cables is proposed with the aid of the optimized CNNs. The feature extraction and recognition are two major phases deployed in this research work. The data collected by VSB from the power cable is subjected for proposed PCA based dimensionality reduction. Then, from these signals, the ROC, RSI, AMA and standard deviation based technical indicator features were extracted. The features of the original power line data were extracted in addition and all these features were together fed as input to optimized CNN. In the CNN, the weight and activation function were optimized via CS-SOA. This evaluation was done by varying the training rate and the noise level. Here, the performance of the proposed model is learned for every variation in the training rate (TP), say 60%, 70%, 80% and 90%, respectively. The TP evaluation is done to comprehend the performance level improvement of the presented work over the traditional works like SVM, CNN and LSTM. Then, the performance of the presented work is compared over the existing works in terms of noise level. Here, during the testing phase, a white Gaussian noise is implied to the collected original signal and the performance of proposed work is analysed under different measures. This evaluation is done by varying the noise from 0.1, 0.2, 0.3, 0.4 and 0.5, respectively. Abstract: In the insulating material, the event of PDs takes place due to the air pockets, voids, or imperfections in such fields. In HV components like cables, the PD detection aids in ascertaining the insulating material condition. Consequently, the territory of PD estimation and conclusion is acknowledged as one of the most significant non-destructive methods for surveying the quality and specialized respectability of HV power mechanical assembly and links. In this work, a new data-driven model to recognize the condition of PD pulses of power cables is proposed with the aid of optimized Convolutional Neural Networks (CNNs). The two main stages of the developed framework are feature extraction and recognition. The data downloaded by VSB from the power cable is subjected to dimensionality reduction via Principal Component Analysis (PCA). Then, from these signals, the features are extracted with the aid of technical indicators like Rate of change (ROC), Relative Strength Index (RSI), Adaptive Moving Average (AMA), and Standard Deviation. Additionally, the features of the original power line data are also extracted. These features are put into an enhanced CNN as input. A new hybrid version known as the Combined Sealion-Swarm Optimization Algorithm (CS-SOA) optimizes the weight and activating function of CNN to increase the accuracy of the PD condition observation. From the result, it can be noticed that the accuracy of the proposed model, is 77.7%, 5.55%, and 3.335% higher than the existing models like SVM, CNN, and LSTM at TP= 80.