Consistency-based diagnosis relies on the computation of discrepancies between model predictions and sensor observations. The traditional assumption that these discrepancies can be detected accurately (by means of thresholding for example) is in many cases reasonable and leads to strong performance. However, in situations of substantial uncertainty (due, for example, to sensor noise or model abstraction), more robust schemes need to be designed to make a binary decision on whether predictions are consistent with observations or not. However, if an accurate binary decision is not made, there are risks of occurrence of false alarms and missed alarms. Moreover when multiple sensors (with differing sensing properties) are available the degree of match between predictions and observations of each sensor can be used to guide the search for fault candidates (selecting candidates “closer” to sensor observations that are more likely to be inconsistent with corresponding predictions). Using Bayesian networks, we present in this paper a novel approach to candidate generation in consistency-based diagnosis. In our formulation, automatically generated Bayesian networks are used to encode a probabilistic measure of fit between predictions and observations. A Bayesian network inference algorithm is used to compute most probable fault candidates taking into account the degree of fit between predictions and observations for each individual sensor.