Partial Discharge Recognition with a Multi-Resolution Convolutional Neural Network

Abstract

Partial discharge (PD) is not only an important symptom for monitoring the imperfections in the insulation system of a gas-insulated switchgear (GIS), but also the factor that accelerates the degradation. At present, monitoring ultra-high-frequency (UHF) signals induced by PDs is regarded as one of the most effective approaches for assessing the insulation severity and classifying the PDs. Therefore, in this paper, a deep learning-based PD classification algorithm is proposed and realized with a multi-column convolutional neural network (CNN) that incorporates UHF spectra of multiple resolutions. First, three subnetworks, as characterized by their specified designed temporal filters, frequency filters, and texture filters, are organized and then intergraded by a fully-connected neural network. Then, a long short-term memory (LSTM) network is utilized for fusing the embedded multi-sensor information. Furthermore, to alleviate the risk of overfitting, a transfer learning approach inspired by manifold learning is also present for model training. To demonstrate, 13 modes of defects considering both the defect types and their relative positions were well designed for a simulated GIS tank. A detailed analysis of the performance reveals the clear superiority of the proposed method, compared to18 typical baselines. Several advanced visualization techniques are also implemented to explore the possible qualitative interpretations of the learned features. Finally, a unified framework based on matrix projection is discussed to provide a possible explanation for the effectiveness of the architecture.

Multi-Scale Analysis and Pattern Recognition of Ultrasonic Signals of PD in a Liquid/Solid Composite of an Oil-Filled Terminal

“…Multi-resolution analysis is often referred to as multi-scale analysis [19][20][21]. This method is a theory based on the concept of function space, and the idea comes from practical engineering.…”

Partial Discharge Classification Using Deep Learning Methods—Survey of Recent Progress

“…They obtained a 100% performance with the CNN method and they compared it with an SVM classifier, where input features were calculated by advanced signal processing methods such as Hilbert-Huang Transform (91.7% accuracy) and Wavelet Transform (96.7% accuracy). Years later, the same authors published another CNN architecture combined with a LSTM network [24]. In this case, for each UHF signal they calculate three different STFTs considering different window lengths to represent the signal in three types of spectrograms: (a) high time resolution, (b) high frequency resolution and (c) medium resolution.…”

“…The structure proposed is shown in Figure 10. Years later, the same authors published another CNN architecture combined with a LSTM network [24]. In this case, for each UHF signal they calculate three different STFTs considering different window lengths to represent the signal in three types of spectrograms: (a) high time resolution, (b) high frequency resolution and (c) medium resolution.…”

Partial Discharge Pattern Recognition of Gas-Insulated Switchgear via a Light-Scale Convolutional Neural Network

“…To effectively solve this problem, deep learning methods that rely on automatic feature extraction are introduced into GIS PD pattern recognition. At present, these deep learning models include LeNet5, AlexNet, one-dimensional convolution, and long short-term memory (LSTM) models [30][31][32][33][34].…”