Interpretable Fault Diagnosis for Overhead Lines with Covered Conductors: A Physics-Informed Deep Learning Approach

Protection and Control of Modern Power Systems, vol. 10, no. 2, pp. 25-39, March 2025, doi.org/10.23919/PCMP.2023.000159

Interpretable Fault Diagnosis for Overhead Lines with Covered Conductors: A Physics-Informed Deep Learning Approach

Genghong Lu¹, Chi Wai Tsang², Ho Nam Yim², Chao Lei¹, Siqi Bu³, Winco K. C. Yung⁴, and Michael G Pecht⁵
¹ Department of Electrical and Electronic Engineering, Centre for Advances in Reliability and Safety, The Hong Kong Polytechnic University, Hong Kong, China
² Centre for Advances in Reliability and Safety, Hong Kong, China
³ Department of Electrical and Electronic Engineering, Research Centre for Grid Modernisation, International Centre of Urban Energy Nexus, Centre for Advances in Reliability and Safety, Shenzhen Research Institute, Research Institute for Smart Energy, Policy Research Centre for Innovation and Technology, The Hong Kong Polytechnic University, Hong Kong, China
⁴ Department of Industrial and System Engineering, Centre for Advances in Reliability and Safety, The Hong Kong Polytechnic University, Hong Kong, China
⁵ Center for Advanced Life Cycle Engineering, University of Maryland, College Park, MD, USA

For more information about this article and related research, please contact Prof. Michael Pecht.

Abstract:

Partial discharge (PD) activity is an indicator of insulation deterioration and by extension, the reliability of power lines. Existing data-driven methods, while helpful, treat PD detection as a binary classification problem, thereby failing to provide physical information (e.g., filter PD pulse), and often provide results that contradict physical knowledge. To tackle this challenge, this paper develops a physics-informed temporal convolutional network (PITCN) for PD diagnosis (i.e., PD detection and PD pulse filtering). During training, physical knowledge of the background noise and PD pulse identification is integrated into a learning model. Once the model is trained, the PITCN can automatically detect PD activity from time-series voltage signals with different background noises and filter PD pulses. Experimental results demonstrate that the developed PITCN outper-forms the rest of the data-driven methods implemented, and in particular, the Matthews correlation coefficient of PITCN surpasses the conventional temporal convolutional network by 0.21.

This article is available for free online here.