Internal short circuit faults are among the most common causes of unexpected transformer failure, and infrared (IR) thermography gives a safe, non-contact way to catch such faults early. This paper presents a complete deep learning pipeline for thermal image based fault severity diagnosis, built and tested on a public benchmark of 255 IR images [15] of a 1 phase dry type transformer covering nine severity classes (Healthy plus eight levels of short circuited winding turns). The pipeline combines physics based thermal feature extraction, a compact convolutional neural network (CNN) classifier verified with Grad CAM, two label free severity estimators (a healthy only auto encoder and K means clustering), and a closed form severity index that needs no training at all. Because the dataset is small, the paper's main contribution is a leakage safe, three condition data augmentation stress test that compares no augmentation, on the fly augmentation, and a fixed offline 10× augmented set. All three models score a perfect 100% on a clean test set, but under a deliberately distorted 'stress test' their accuracy is 43.6%, 71.8%, and 97.4% respectively showing that clean test accuracy alone hides a large robustness gap. Physics features correlate strongly with true severity (Spearman ρ up to 0.983), and the label free estimators reach ρ = 0.987–0.989, nearly matching the supervised CNN while needing little or no fault labelled data. These results give utilities a practical, layered path to deploy thermal fault diagnosis even when labelled fault data is scarce.

Transformer fault diagnosis, infrared thermography, convolutional neural network, Grad CAM, data augmentation, robustness, auto encoder, K means clustering, severity index.