In this paper, for the fault detection of a single-input single-output (SISO) system, we define a concept of matching chaotic stimulation with a measured system, establish a criterion of matching chaotic stimulation with a measured system, and propose a method of adjusting the matching of chaotic stimulation with a measured system. Finally, the improved prediction error is used to detect faults in the SISO system. The simulation results show that matched chaotic stimulation can be used to detect faults effectively, while unmatched ones can add uncertainty into the fault detection.