A bi-level machine learning method for fault diagnosis of oil-immersed transformers with feature explainability

Power transformer faults are considered rare events, so data samples in normal operations are much more readily available than in faulty conditions. Traditionally, power transformer fault diagnoses were enabled through gas-in-oil data, where erroneous diagnoses of faulty conditions as normal could have a more significant effect on power system operations than wrong diagnoses of normal operations as a faulty condition. Therefore, it is imperative to analyze gas-in-oil data characteristics more effectively to improve the performance of diagnostic methods. In this paper, an explainable bi-level machine learning method is proposed for oil-immersed power transformer fault diagnoses, consisting of a binary imbalanced classification model and a multi-classification model. The proposed Extreme Gradient Boosting models are designed with custom functions at each level, and automatic hyperparameters tuning is conducted based on Bayesian optimization. A fault feature selection is developed using the SHapley Additive exPlanations method to explain the diagnosis results, which could mine the impacts of fault features on diagnosis results and find the approach to improve the model performance. The fault diagnosis results are presented with performance analysis and comparative studies, and the feature selection results with importance analysis for each fault type based on SHAP value is provided, which demonstrates the feasibility and effectiveness of the proposed method.