Development of a Machine Learning-Based Fault Detection System for Industrial Power Transformers in Rivers State

Background: Transformer failures in Rivers State industrial facilities cause costly production outages and safety hazards. Traditional dissolved gas analysis is periodic and labour-intensive, while machine learning-based continuous monitoring offers significant predictive maintenance advantages. Aim: This study developed a machine learning-based fault detection system for industrial power transformers in Rivers State, South South Nigeria. Methods: A dataset of 1,240 dissolved gas analysis records from 85 transformers over five years was assembled from industrial facility maintenance records. Feature engineering extracted nine fault-relevant gas concentration ratios. A random forest classifier, support vector machine, and gradient boosting ensemble were trained and compared using 5-fold stratified cross-validation. Results: The gradient boosting ensemble achieved the highest accuracy of 94.2% and F1-score of 0.93 in fault type classification across five IEC 60599-defined fault categories. The model successfully identified 11 of 12 incipient faults in a prospective validation set six weeks before conventional detection. Conclusion: Machine learning-based transformer fault detection substantially outperforms traditional rule-based methods and enables early intervention. Integration with online gas-in-oil monitoring sensors for real-time fault prediction is recommended for critical industrial transformers in South South Nigeria. Keywords: transformer fault detection, machine learning, dissolved gas analysis, Rivers State, predictive maintenance.