In this thesis we propose a general framework for distributed diagnosis. Each diagnosis instance consists of two phases: local estimation, and inter-component communication for consistency. For the latter phase we introduce the concepts of supremal global support (for global consistency) and supremal local support (for local consistency). We provide a computational procedure CPGC for achieving supremal global support, and CPLC for supremal local support. The two supremal supports lead to distinct distributed diagnosis problems. It turns out that supremal global support results in better quality of diagnosis in the sense that fewer fault candidates are reported in each diagnosis instance; but supremal local support results in a computational procedure that is better scalable as long as it can terminate. In practice the two supremal supports may be combined for a satisfactory tradeoff between quality of diagnosis and scalability of the diagnoser. To reduce time complexity of CPGC, we propose a hierarchical computational procedure, utilizing multi-resolution diagnosis. Although high-level abstract models for hierarchical computation need extra memory, our numerical results show that the overall space complexity as measured by memory usage in storing both the models and the intermediate computational results is no worse (and in some cases better) than the space complexity in our non-hierarchical approaches. Finally, we explain how to use probabilistic reasoning to reduce diagnostic ambiguity without inserting extra sensors.