This paper describes the development of a computational tool based on the application of probabilistic inference networks (Bayesian networks) to deal with the uncertainties related to the transformer diagnosis problem, considering information from alarms. A Bayesian network is a graphical representation of a problem where each node is an event (alarm or fault) and the links represent causal relations between them. The rates of correct and incorrect operation of relays, as well as their failure to operate when necessary, are essential to the implementation of this tool. These rates were calculated based on a historical database of a transmission company and are also discussed in the paper. The association of a diagnosis with a probability facilitates the interpretation of the program results, reducing the risk of an operator making a wrong decision during stressful situations.