Three-Dimensional Modeling of Multi-Source Heterogeneous Data for Power Substation Equipment

This paper presents a comprehensive study on multi-modal imaging techniques for fault diagnosis and condition assessment of critical power equipment, with a focus on transformer bushings. The research integrates visible light imaging, thermal infrared imaging, ultraviolet (UV) imaging, and acoustic imaging to detect and analyze defects such as insulation degradation, overheating, partial discharge, and mechanical vibrations. A novel fault diagnosis framework is proposed, combining offline model training and online diagnosis. The framework employs advanced algorithms, including Mask R-CNN for target detection, an improved Pulse-coupled Neural Network (PCNN) with Simple Linear Iterative Clustering (SLIC) for fault region extraction, and Extremely Randomized Trees for feature selection. Additionally, Generative Adversarial Networks (CycleGAN) are utilized to augment limited fault datasets through synthetic sample generation. Experimental validation on a 110 kV oil-paper bushing demonstrates the effectiveness of the proposed approach in enhancing fault detection accuracy and diagnostic reliability. The study highlights the potential of multi-modal imaging and machine learning in improving the maintenance and operational safety of power infrastructure.