Transformer fault analysis based on improved DS evidence theory and Bayesian network

To quantify the risk of transformer faults, the improved D-S evidence theory combined with the Bayesian network is proposed to address the effects of missing samples, misinformation, and omission of information on fault diagnosis accuracy. The transformer vibration spectrum dataset under different electrical working conditions is downsized using principal component analysis (PCA), and the conflict degree values are assigned by weights through the improved Dempster’s synthesis rule to obtain the probability assignment function of each fault risk level that integrates experts’ opinions. Based on the logical relationship between nodes, the Gaussian affiliation function is constructed, and the fuzzy probability of fault risk is determined by combining expert experience. By introducing the vibration spectrum characteristic values of transformer operation into the DS-BN model for iteration, the forward and reverse inference of the Bayesian network is utilized to realize the fault diagnosis of the transformer, which shows significant feasibility and improves the accuracy of fault diagnosis.