A Cycle Consistent Dual-Domain Transformer Framework for Detection of Partial Discharge

Over the past decade, scholarly inquiries have delved into the critical issue of insulation failure attributed to partial discharge (PD) phenomena in overhead transmission lines with insulated conductors, as well as in electrical apparatus such as generators and motors across various sectors. The development of a proficient PD detection framework could result in considerable maintenance savings and avert electricity supply interruptions. Conventional techniques depend on manually curated features and sector-specific knowledge for the identification of PD patterns within electrical currents. Recent advancements have seen a surge in the proposition of data-driven, deep learning methodologies aimed at obviating the need for these bespoke feature extractions. Nonetheless, these innovative approaches predominantly analyze data either in the time or frequency domain. A significant body of research has been dedicated to producing phase-resolved partial discharge (PRPD) patterns from unprocessed PD sensor data, extracting pivotal attributes of these patterns to furnish a visual diagnostic tool for exhaustive defect analysis. PRPD diagrams infer a linkage between PD events within the positive and negative half-cycles of an alternating electrical waveform. This inter-cycle correlation, however, has not been thoroughly investigated in deep learning strategies. The present study introduces an innovative feature fusion model, termed Dual-Trans Net, designed to harness time, frequency, and phase domain features for unified learning within a singular framework. An empirical evaluation of our framework, employing a publicly available dataset from high-frequency voltage sensors for the identification of compromised power lines, has culminated in a leading-edge Matthews Correlation Coefficient (MCC) score of 0.7830.