Vibration signal of rotor system in fault state tends to be highly non-linear， which is mainly manifested in the frequency domain as coupling between different frequencies， generating combined frequency， differential frequency and other combined frequencies. In order to solve the problem that traditional spectrum analysis only focuses on the frequency component and its magnitude of the signal， but ignores the phase information of the signal. Bispectrum contains signal phase information and is sensitive to nonlinearity. It can amplify weak nonlinearity of early faults and detect nonlinear phase coupling between different frequencies in the spectrum. By implanting different types of faults into ZT-3 rotor test-bed， the acceleration signals of the system in different states are collected， the information entropy of each frequency band is extracted from the bispectrum of vibration signals， and the fuzzy clustering method is used for fault identification. The result shows that the bispectrum entropy as a characteristic parameter can accurately identify the fault type of the rotor system， which verifies the feasibility of the method.