Abstract In order to diagnosis vibration fault of centrifugal pump accurately, aiming at the non-stationary characteristics of the vibration signals of centrifugal pump, a fault diagnosis method based on recurrence quantification analysis was put forward. First of all, the recurrence quantification analysis( RQA) method was used to extracted nonlinear characteristic parameter of the vibration signals, and the feature vector was generated by RQA nonlinear characteristic parameter. The feature vectors were employed as the input samples to train a modified Elman neural network, and then the running state classifier of the centrifugal pump was set up. The experimental results show that proposed method is effective for centrifugal pump fault diagnosis.