Abstract: Aiming at the non-stationarity of rotor fault signal and the inability to effectively extract sensitive fault features, a fault diagnosis method was proposed by combining the Volterra model of variational mode decomposition (VMD) and singular value entropy. The parameter selection methods affecting the accuracy of VMD decomposition were deeply studied, and the solutions to the related problems were given. Firstly, the measured rotor signals under different working conditions were decomposed by VMD, and the intrinsic mode function (IMF) sensitive to fault characteristics was selected by using the increment of energy entropy for phase space reconstruction, so as to establish the Volterra adaptive prediction model, and the model parameters were used as the initial eigenvector matrix. Then, the initial eigenvector was decomposed by singular value decomposition to obtain singular value entropy and singular value eigenvector matrix, which were used to describe the fault characteristics of rotor. Finally, the fuzzy C-means (FCM) algorithm was used to identify the rotor working state and fault type. The experimental results show that the proposed method can effectively realize the feature extraction and type recognition of rotor fault. Compared with ensemble empirical mode decomposition (EEMD), it is proved that this method has more effective fault feature extraction performance and is a feasible method.