The fault hierarchy representation is widely used in expert systems for the diagnosis of complex mechanical devices. On the assumption that an appropriate bias for a knowledge representation language is also an appropriate bias for learning in this domain, we have developed a theory revision method that operates directly on a fault hierarchy. This domain presents several challenges: A typical training instance is missing most feature values, and the pattern of missing features is significant, rather than merely an effect of noise. Moreover, features that are present may be corrupted. Finally, the accuracy of a candidate theory is measured by considering both the sequence of tests required to arrive at a diagnosis, and its agreement with the diagnostic endpoints provided by an expert. We describe the DELTA algorithm for theory revision of fault hierarchies, its application in knowledge base maintenance, and report on experiments with a diagnostic system in current use.