With the improvement of control system composition and operation process complexity, the uncertainty in its operation process increases and real-time observation data is difficult to obtain, and the influence of noise also exists in the process of signal acquisition, which brings more difficulties to the prediction of the remaining useful life (RUL). To solve these problems, a hybrid multi-stage methodology for RUL prediction of control system is proposed. The variant of unscented Kalman filter (UKF) utilizes dynamic Bayesian networks (DBNs) for uncertainty analysis in the process of prediction using UKF, to analyze RUL of nonlinear degenerate systems. In the prophase of prediction, the dynamic unscented Kalman filter models calculate the distribution of random faults and process noise, match the degradation stage of the system and obtain the operation data. Then, optimize the degradation process of the system, and the covariance and the optimal estimate of the system are calculated by cyclic iteration. The real degradation process of control system is simulated by optimizing the results, so as to compensate for the lack of accurate measurement of the real degradation process. The proposed method can improve the accuracy of RUL prediction and enhance the robustness of the prediction model. The methodology is verified by subsea Christmas tree with electro-hydraulic compound control. © 2022 Elsevier Ltd