Optimizing Thermal Image Classification and Fault Detection for Electrical Distribution Transformer Room components using ResNet and VGG16

This research investigates the use of deep learning models, specifically modified versions of VGG16 and ResNet50 architectures, for thermal image classification in electrical transformer rooms. Overheating and related faults in power transformers pose significant risks to the efficiency, safety, and longevity of electrical distribution systems. The study utilized a dataset of 3,851 thermal images classified into six distinct categories, representing various normal and faulty conditions in transformer room components. The models were trained using five-fold cross-validation, and data augmentation techniques were applied to enhance model robustness. The modified ResNet50 achieved an overall accuracy of 76.89%, outperforming the modified VGG16, which achieved 73.07%. Both models demonstrated strong potential in automating the detection of faults, thereby minimizing human intervention and improving maintenance efficiency. The findings underscore the applicability of transfer learning and CNNs in enhancing the reliability of electrical systems, The ResNet50 and VGG16 models were chosen and focused on due to their strong performance on image classification tasks, especially for smaller datasets such as the data used in this proposed research. Its proven architectures make it a suitable starting point for this type of work.