Deep learning for partial discharge diagnosis in electrical assets: fundamentals, applications, and future perspectives

Abstract Partial discharge (PD) monitoring and detection are crucial for ensuring the longevity and reliability of high-voltage apparatus. PD signals acquired in a laboratory setup or from on-site measurements are often complex, noisy, and high-dimensional. Therefore, effective signal processing is necessary to extract meaningful information for an accurate diagnosis. Deep learning (DL) algorithms have gained significant attention in recent years for analyzing PD patterns because they can learn complex representations without requiring manual feature extraction. Current research on DL for PD diagnosis is fragmented, employing various pre-processing techniques and architectures. Consequently, a review can illuminate successful strategies and areas necessitating enhancement. This review comprehensively examines the application of various DL architectures in PD analysis and their effectiveness on different electrical equipment. Details of the challenges the research community faces are discussed, along with practical considerations. This review also explores, in detail, the differences, limitations, and steps to be considered, as well as the validation strategies between laboratory and field testing. Finally, the review highlights emerging research directions and presents a framework to support future developments. This article serves as a roadmap for applying advanced computational intelligence techniques in PD diagnosis and highlights the potential of DL to support scalable condition monitoring of high-voltage assets.