Abstract: The ability of Bayesian network to deal with uncertain problems can be a solution to transformer fault diagnosis when data is incomplete and reliable conclusions are difficult to be reached. By combining the Bayesian network classifier and rough set reduction theory, this paper set up a Bayesian network classification model based on expert knowledge and statistical data. The integral use of DGA and electrical test data as the input set of diagnosis realized the probabilistic reasoning and sequencing of the possible failure types, and as a result, improved the reliability of the diagnosis. Meanwhile the rough set reduction theory was used to make the minimization reduction of Bayesian network classification model, effectively reduced the complexity of network structure and the input of the model. Experiments proved that this method is an effective transformer fault-diagnosis method and has the ability to deal with the absence of information, the fault-tolerant features and high accuracy.

Key words: transformer, failure diagnosis, decision table, Bayesian network, rough sets, information reduction