A method for fault detection in industrial systems is presented. Plant devices, sensors, actuators and diagnostic tests are described as stochastic Finite State Machines. A formal composition rule of these elementary models is given to obtain: (a) the set of admissible fault signatures, (b) their conditional probability given any fault event, (c) the conditional probability of a fault given a prescribed signature. The modularity and flexibility of this approach make it suitable to deal with complex systems made by a large number of elementary models.