A Foundational Framework for Substation Fault Diagnosis from Imbalanced Thermal Data Using Transfer Learning

Abstract

A key component of predictive maintenance for vital electrical substation equipment is thermographic inspection, which makes it possible to identify thermal abnormalities early on before they cause failures. Maintenance workflows suffer by the subjective, time-consuming, and human error-prone manual interpretation of the resulting thermal images. Although deep learning offers an efficient automation solution, a significant real-world obstacle to its widespread use is the extreme scarcity and class imbalance of available fault data. In order to close this gap, a comprehensive methodological framework for creating a reliable baseline diagnostic model is proposed and described in this paper. The strategy is based on transfer learning, which involves optimizing strong, pre-trained Convolutional Neural Network (CNN) architectures to take advantage of their acquired features and lessen the impact of sparse data. From data preparation and aggressive minority fault class augmentation to the use of a weighted loss function during training, the framework describes the complete pipeline. The experimental pipeline was validated after this methodology was put into practice. The model's anticipated initial bias towards the prevalent "Normal" class—a direct result of the data imbalance—was validated by preliminary observations. The main result of this work is this procedural validation, which provides a strong, repeatable basis for further investigation. While conclusive performance analysis and the investigation of explainability techniques remain open for future work, this study offers a practical route for creating trustworthy diagnostic tools under practical data constraints.