Enhancing Predictive Maintenance of Power Transformers Through Machine Learning Approaches

This paper focuses on the use of machine learning algorithms to assist in predictive maintenance, aiming to reduce downtime and associated costs for electrical equipment. It introduces two machine learning predictive indicators: the Chromatographic Assay Indicator (CAI) and the Electrical Failure Risk Indicator (EFRI), which leverage chromatographic and sensor data, respectively. The CAI indicator was trained on external data, and its generalization capability was assessed using company-specific chromatographic data with expert assistance. The evaluation involved two heuristics, both focusing on keyword identification in free-text diagnoses. Results showed over 90% accuracy in predicting failures in reactors and power transformers, and an AUC of nearly 85% for reactors. Additionally, CAI outperformed classical Dissolved Gas Analysis (DGA) methods. On the other hand, the EFRI indicator was trained using company’s internal data from a monitoring system of operational sensors and maintaining power transformer data. The classification model achieved over 95% accuracy and an AUC of almost 90% on the test set. These results indicate that both indicators can be integrated into a solution to support maintenance specialists in their decision-making processes.