Subsystem Measurement-Based Condition Assessment for Power Transformers via Joint Inference of Data and Knowledge

Based on various measurements, the condition assessment of power transformers ensures transformer reliability for achieving a stable power supply and improves economic benefits. However, the existing methods suffer from artificially summarized rules, heavy data dependence, incomprehensive measurement analysis, superficial information fusion, or unexplainable assessment results. Therefore, an interpretable joint inference method based on subsystem measurements is proposed to automatically extract the knowledge and data features, enabling profound information fusion and reducing the impact of poor-quality data. First, the sparse autoencoders (SAE) and graph convolutional network (GCN) extract features from the subsystem condition data and the archive knowledge graph representing the maintenance histories, respectively. Next, the method assigns weights to the features according to the mask vectors. Then, a Bayesian neural network (BNN) analyzes uncertainties to recognize the condition grade, and the health index is calculated through fitting distributions and Monte Carlo sampling. Finally, a local interpretable model is designed to interpret the decisions made by the proposed method. Verified by experiments, the proposed method achieves an F-measure of 97.28% on grade recognition, which is 20.94% higher than contrast models. Moreover, the method is also proven to outperform the other models in dealing with disturbed and incomplete data.