This paper shows how a Bayes classifier can be implemented for a failure detection system where statistical failure data is not available for one of the classes. Results of field data obtained from a large electric power generator are shown. The classifier is further improved by the iterative re-evaluation of the prior probabilities, which results in the use of higher alarm threshold values when a good agreement between the monitored quantity and its estimated value is observed, while large disagreement values result in smaller thresholds. As expected, the proposed system is an improvement over a classical Bayesian implementation and a large improvement over a fixed, arbitrary value threshold classifier.