AI BASED FAULT DIAGNOSIS IN POWER TRANSFORMER-A DEEP LEARNING FRAMEWORK USING CONVOLUTIONAL NEURAL NETWORK BASED ON INFRARED IMAGING AND WASSERSTEIN GENERATIVE ADVERSARIAL NETWORK TOOLS

ABSTRACT

The condition of a transformer impacts the safety of electrical power networks. Transformer failures not only reduce system reliability but also cause serious accidents and financial losses. Various diagnostic techniques exist to monitor transformer operation, but most are challenging for real-time monitoring due to difficulty in online identification, noise interference, and high maintenance costs. This article proposes an overheating defect diagnosis method using forward-looking infrared (FLIR) thermography images and introduces a deep learning framework for power transformers. The proposed thermal camera system collects four types of FLIR images. Wasserstein Autoencoder Reconstruction (WAR) and Convolutional Neural Networks (CNN) models have been used to process these images. The trained CNN and WAR models have been combined in series to form a defect diagnosis module. The aim is to develop a cost-effective monitoring approach using remotely captured thermal images of a power transformer in an outdoor substation. The CNN model has been used to process these images to extract detailed feature parameters, preventing incomplete information issues from various feature-extraction techniques. Transformer health is classified into four categories: Monitor, Likely Defect, Repair at the Next Opportunity, and Major Defect Repair Immediately, based on temperature rise above ambient. The combined WAR–CNN technique has been used to improve remote monitoring. This approach has resulted in identifying common defects with high accuracy and has offered significant benefits over state-of-the-art methods, compact storage needs, and high diagnostic accuracy.