When designing model-based fault-diagnosis systems, the use of consistency relations also called e.g. parity relations is a common choice. Different subsets are sensitive to different subsets of faults, and thereby isolation can be achieved. This paper presents an algorithm for finding a small set of submodels that can be used to derive consistency relations with highest possible diagnosis capability. The algorithm handles differential algebraic models and is based on graph theoretical reasoning about structure of the model. An important step, towards finding these submodels and therefore also towards finding consistency relations, is to find all minimal structurally singular (MSS) sets of equations. These sets characterize the fault diagnosability. The algorithm is successfully applied to a large nonlinear industrial example, a paper plant.