Partial Discharge Detection with Convolutional Neural Networks

Abstract

Covered conductors are widely adopted in the medium to low voltage systems to prevent faults and ignitions from events such vegetation contacting with distribution lines and conductors slapping together. However, such events could cause partial discharge in deteriorated insulation system of covered conductors and ultimately lead to failure and ignition. To prevent power outages and wildfires, it is crucial to detect partial discharges of overhead power lines and perform predictive maintenance. In this paper, we develop advanced machine learning algorithms to detect partial discharge by using measurements from high frequency voltage sensors. Our innovative approach synergistically combines the merits of spectrogram feature extraction and deep convolutional neural networks. The proposed algorithms are validated using real-world noisy voltage measurements. Compared to the benchmark, our approach achieves notably better performance. Furthermore, the classification results from the neural networks are interpreted with an occlusion map, which identifies suspicious time intervals when partial discharges occur.