Development and Performance Evaluation of Machine Learning Model for Fault Diagnosis of Mineral Oil-Filled Transformer Using No-Code Approach

Abstract

Interpretation of dissolved gas composition is crucial in fault diagnosis of oil-filled transformer.  Conventional interpretation methods have limitations in terms of accuracy and adaptability to new fault patterns, while artificial intelligence techniques applied in this field require extensive programming knowledge, data volumes and computational resources. This study presents the development and performance evaluation of machine learning models for transformer fault diagnosis using a no-code approach. Utilizing Microsoft Azure Machine Learning, trained on 639 ground truth datasets, the Gradient Boosting model consistently outperformed other models and Duval Triangle, achieving the highest accuracy of 81.25% with 2 false negatives. Experiments were conducted to optimize the model's performance, including varying train-test ratios, applying oversampling and outlier treatment.  The model was validated through field trials on 4 transformers, successfully predicting faults for all transformers, consistent with forensic findings. Feature importance analysis results were consitent with established dissolved gas–fault relationships, which provided transparency into the model’s fault classification and supported user confidence. This research highlighted the practical applicability and robustness of the machine learning model, offering a singular diagnostic method for field engineers to distinguish between healthy and faulty transformer, and accurately identify the fault type and severity. Leveraging the no-code approach,  field engineers, being the subject matter experts, could harvest the advantages of artificial intelligence and impart their knowledge and data, without vast computing knowledge and resources.