An interpretable machine learning method for fault diagnosis of oil-immersed transformers based on edge inference

• A filtered feature selection algorithm based on real domain rough set theory is proposed. • Deployed the optimal model at the edge-end. • The proposed model is made interpretable using the SHAP method. • The influence of characteristic gases on transformer fault types in model decision making is discussed through actual cases. Intelligent diagnostic models using dissolved gas analysis are crucial for oil-immersed transformer fault diagnosis. However, the inherent “black box” nature of these models limits interpretability, and traditional methods that upload local data to central servers raise data security concerns. To address these issues, this study proposes an interpretable fault diagnosis model for edge deployment. First, a filtered feature extraction algorithm based on real domain rough set theory is proposed to optimize feature extraction before model input. Experimental results demonstrate that this algorithm enhances model performance and reduces inference time at the edge-end. Second, the hyperparameters of Extreme Gradient Boosting are automatically tuned using the Newton–Raphson optimizer. Compared with other diagnostic methods, the proposed model yields superior classification effect accuracy. Following edge-end inference, the SHapley Additive exPlanations method is employed to analyze feature impact on diagnostic results, visualizing the significance of different characteristic gases for fault types using SHAP values. Finally, the model’s robustness, reliability, and interpretability are validated through real cases, providing practical insights for transformer operation and maintenance.