Learning diagnosis models using variable-fidelity component model libraries

System models that are used in model-based diagnosis are often composed of components drawn from component libraries. In these component libraries, there may be multiple systems of equations per component (component implementations). For example, a component may be modeled as a non-linear system (high-fidelity model), linear system, and a qualitative system (low-fidelity model). Choosing the right component model for system diagnosis is a difficult task and requires a search in the space of all possible component type combinations. In this paper we propose a method that automates this task and computes a system model that optimizes a set of diagnostic metrics in a set of diagnostic scenarios. Initial experimental results show that having linear models of some of the components in a system preserves the diagnostic accuracy and isolation time while, at the same time, improves the computational complexity and numerical stability.