This work aims to study which sensors are required to be installed in a process in order to improve certain fault diagnosis specifications. Especially, the present method is based on structural models. Thus, system models involving a wide variety of equations (e.g. linear, non-linear algebraic, dynamics) can be easy handled. The use of structural models permits to define the diagnosis properties from the Dulmage-Mendelsohn decomposition, avoiding in this way the computation of any minimal redundant subsystem. Furthermore, in the present paper, the cost of the sensor configuration is considered. Therefore the proposed method attempts to find not all the possible solution but the optimal one. The optimal search is efficiently performed by developing an algorithm based on heuristic rules which, in general, allow to significantly reduce the search.