A Partial Discharge Detection Method for High-Voltage Cables Based on CycleGAN and Transfer Learning

Detecting partial discharge (PD) in high-voltage cables is an effective method for assessing cable insulation. Traditional PD detection methods often struggle in real-world environments due to strong noise interference. Deep learning presents a promising direction for PD detection, but it typically necessitates paired datasets where each collected noisy PD signal needs a corresponding pure PD signal. However, obtaining such paired signals is impossible in real-world scenarios.In this paper, we propose a novel PD detection method based on CycleGAN and transfer learning. Using CycleGAN can address the challenge that the pure PD signals corresponding to noisy PD signals are unavailable, which makes it well-suited for real-world applications. Moreover, we employ a transfer learning strategy, utilizing abundant numerically simulated data to develop a high-quality pre-trained model. Subsequently, based on this pre-trained model, we can achieve a well-performing model by fine-tuning it with the limited laboratory data based on physical setup. We also design a comprehensive loss function that builds upon the loss of CycleGAN by incorporating a new item, i.e. soft dynamic time warping (Soft-DTW) loss. This Soft-DTW loss effectively evaluates the similarity between signals, even when they are distorted or twisted, thus enhancing the performance of PD signal detection. Extensive experiments on both the simulation dataset and laboratory dataset validate the effectiveness of our method.