Enhanced Visual Identification of Dimensionally Expanded FDS for Transformers Insulation Aging State Assessment

Accurate identification of the aging state of insulation is crucial for preventing potentially catastrophic failures in transformers. Dielectric frequency domain spectroscopy (FDS) has shown tremendous potential for on-site diagnostic testing of insulation performance in the current scenario. However, precise evaluation of the aging state requires extraction of age-sensitive features from the relaxation information stored in FDS. Therefore, a novel deep-learning framework is proposed in this paper leveraging Gramian angular summation fields (GASF), Gramian angular difference fields (GADF), and Markov transition fields (MTF) feature parameters derived from transformed FDS data. It captures intricate temporal and spatial patterns inherent in FDS. Furthermore, a network-based insulation diagnosis model is realized using convolutional neural networks (CNN) based on the dimensionality expansion approach of the FDS, enabling accurate identification of transformer insulation aging state dynamics. The deep learning-based aging diagnosis network has an accuracy of 99.62% in eight aging states. Additionally, the robustness of the aging diagnosis model to moisture variation is also explored, demonstrating the method's ability to identify the spectra' overall features. The present contribution establishes the utility of extracted parameters from relaxation information in FDS as a valuable tool for assessing the aging of the transformer's oil-paper insulation system.