<![CDATA[The growing need for high-quality electrical infrastructure in modern data centers and grid systems has driven the adoption of more effective diagnostic methods to prevent catastrophic failures in high-voltage equipment. Partial discharge (PD), defined as localized electrical discharges that do not completely bridge an insulation gap, is the primary precursor of insulation degradation and equipment failure. Conventional PD detection algorithms, though useful in controlled settings, are constrained in their ability to discriminate noise, recognize patterns under dynamic operational conditions, and make real-time decisions. The incorporation of machine learning (ML) algorithms into PD detection systems has emerged as a transformative technology, enabling unprecedented accuracy in signal classification, anomaly identification, and failure prognostication. This review summarizes recent developments in ML-based PD detection, exploring supervised, unsupervised, and deep learning architectures applicable to offline and online settings. We critically analyze convolutional neural networks (CNNs), recurrent architectures, support vector machines (SVMs), and ensemble approaches to phase-resolved partial discharge (PRPD) pattern recognition, ultra-high-frequency (UHF) signal processing, and acoustic emission analysis. We evaluate implementation obstacles, including dataset limitations, computational costs, model interpretability, and cybersecurity concerns unique to U.S. infrastructure deployment. Competing review papers are identified, and the differentiating scope of the present work is stated explicitly. The review concludes with the identification of essential research gaps and a practitioner-oriented decision framework for developing robust, standardized ML-PD systems suitable for integration into existing predictive maintenance workflows.]]>
Copyrights © 2026
