Diagnostic models are designed specifically to support diagnostic reasoning. We first place diagnostic reasoning within the framework of troubleshooting, the process whereby an incorrectly functioning system is restored to normal function. Diagnostic reasoning determines a set of component faults that can account for observed abnormalities in system function. As such, diagnostic models must incorporate elements of function as well as behavior. We propose that function be introduced in the composition of component behaviors. We discuss two general principles of diagnostic modeling: model variables assume values relative to normal (the "normality principle") and these values are propogated to account for the production of single outputs (the "single output principle"). These two principles yield significant simplifications in the value spaces of model variables and in model structure. We illustrate our approach to diagnostic modeling and reasoning with examples from xerography.