Study On Fault Features And Multi-evidence Information Fusion For Power Transformer Fault Diagnosis

The diagnosis of latent faults of power transformer is the key to implement the condition-based maintenance for large power transformer. But for the existing transformer fault diagnosis method, it is too subjective in the index parameter selection, and through a single method it is difficult to make the diagnosis of transformer faults be comprehensive and objective. Different fault diagnostic methods may lead to inconsistent and even contradictory conclusions. However, because of the emergence of information fusion technology, an effective way can be found to solve the above problems. It has positive significance to conduct the researches on fault features and multi-evidence information fusion diagnosis of power transformer faults, improving the diagnosis level of power transformer faults.① In combination with bathtub curve, the probability characteristics of transformer faults in three different operation stages(early fault stage, accidental fault stage loss fault stage) are discussed. In addition, the reasons for transformer faults in overheating, discharging, short circuit and insulation are deeply explored, and the characteristics parameter sets of major characteristics of four faults are obtained. As major factors which influence insulation performance in different stages are different, the characteristics parameter sets selected for different stages vary.② As the transformer is actually characterized by higher frequency and large dispersion in early fault stage, the original faults diagnosis characteristics parameter sets are determined based on the traditional analysis method of dissolved gas in oil(characteristic gas method, ratio method and graphic method)in combination with electrical test data. On the WEKA platform, through calculating the intensity of characteristic parameter information gains, the fault diagnosis index sets of transformer in early fault stage is established from the oil chromatographic data and a large number of electrical test data.③ Based on BP, RBF neural networks, support vector machines and supervised and self-organizing competitive-type cluster network(S-Kohonen), the primary diagnosis model of transformer is established. If the conclusions drawn through four fault diagnosis methods are consistent, the final diagnosis result can be directly given; otherwise, the evidence body required for the decision-level fusion diagnosis must be constructed to further integrate and inference to obtain the diagnosis conclusions.④ Based on the idea of evidence discount, the multi-evidence body with conflicts are amended in combination with evidence classification discount algorithm, evidence pivot element discount algorithm, solving the problem of combination paradox in the evidence fusion. In addition, the Dempster combination rule is used to establish a fault fusion diagnosis mode of multi-evidence body. The cases of fault diagnosis of two 220 k V transformers and the accuracy of test sample identification show that the fusion diagnosis method of faults proposed in this paper has higher accuracy rate compared to the single fault diagnosis method, which can be used to effectively diagnose the faults of transformer in early operation stage.