Thermography Inspection with Machine Learning for Malfunction Prediction in Power System Equipment

Abstract

This paper presents a machine learning-based approach for early failure detection in power system equipment using thermographic image analysis. While existing studies often rely on complex deep learning models, our method introduces a lightweight yet effective solution combining image preprocessing, unsupervised segmentation, and classical classifiers. The key innovation lies in the integration of automated Region of Interest (RoI) detection and the evaluation of its impact on model performance. A dataset of 624 RGB thermographic images from motors and transformers is used to benchmark various machine learning algorithms, including KNN, SVM, Decision Tree, Random Forest, Naive Bayes, and Neural Networks. Performance metrics such as precision, recall, and F1-score are analyzed and show that RoI segmentation significantly improves classification accuracy (up to 32% for SVM and 15% for Neural Networks). The proposed method is resource-efficient and achieves strong results without requiring temperature vector data. These findings highlight the practical value of the approach for predictive maintenance and early fault diagnosis in industrial environments.