The paper deals with a new active fault diagnosis framework for stochastic large scale systems, that are assumed to be decomposed into weakly coupled input-decentralized sub-systems. The solution to the active detection problem is based on the multiple-model framework and behavior of each subsystem is represented by a set of models describing fault-free and faulty behavior of the subsystems. The centralized, decentralized, and distributed architectures of the active fault detector are introduced, the corresponding optimization problems are set up and approximate solutions to these problems are proposed. The behavior of the detector in the considered architectures is illustrated using a numerical example.