Experimental evaluation of detecting power transformer internal faults using FRA polar plot and texture analysis

Power transformers are prone to short circuit faults and external harsh environmental conditions that may result in various internal winding and core deformations. Frequency response analysis (FRA) is the current technique widely used to detect such kind of faults. The technique relies on scanning each phase in a wide frequency range to detect any discrepancies in the winding FRA signature. Current interpretation process for FRA signatures relies on graphical analysis of the magnitude plot without paying much attention to the phase angle plot. This may result in inconsistent analysis for the FRA signature as so far, there is no standardized and automated technique for FRA signature classification. To overcome this limitation, a new polar plot signature based on the integration of the FRA magnitude and phase angle plots has been proposed in the literature. While this new approach has been validated through simulation analysis, its performance on real transformers has not been assessed yet. This paper is aimed at investigating the practical feasibility of replacing the conventional FRA signature with the polar plot and automate the fault diagnosis process using digital image processing techniques. In this context, various winding faults including short-circuit disks, disk space variation, and axial disk buckling have been implemented on a custom-made, Y/Y three phase, 50 Hz, 400 kVA power transformer. The low voltage winding of the investigated transformer comprises 12-turn helical type while the high voltage winding is made of composite structure comprising 10 disks of interleaved winding on the top, 10 disks winding in the middle and 10 disks of continuous winding in the bottom. Impact of various levels and locations of these faults on the FRA polar plot is investigated through extracting digital image texture analysis features. To ease the classification process, a unique classification metric namely image sum square max-min ratio error is calculated based on the extracted features.