The paper deals with fault detection of stochastic discrete systems with the consideration of unknown disturbance and noises, which are modelled as time-varying parameters and stochastic process, respectively. The disturbance decoupled observer design problem is investigated for the reconstructed singularity system which contains the states and disturbance. Then using hypothesis test technique, the residual generated by the designed observer is analyzed and a set of necessary conditions to the fault detectability is considered. Finally, the proposed fault detection approaches are applied to a practical DC motor to demonstrate the effectiveness of the results.