Adaptive Cooperative Intra–Inter Domain Adversarial Network With Stage Division for RUL Prediction of Ship Electric Propulsion System

The remaining useful life (RUL) prediction of ship electric propulsion system faces severe challenges such as multicondition switching and complex degradation modes. Under complex coupling conditions involving significant distribution differences across multisource domains and quantitative requirements for prediction uncertainties, conventional multisource methods struggle to resolve feature shifts caused by coupled condition-mode differences, as they primarily focus on single-aspect solutions. To this end, an RUL prediction method that integrates the entropic Gromov–Wasserstein (EGW) distance and adaptive cooperative intra–inter domain adversarial network (ACIDAN) is proposed in this article. First, a divergence measurement model for the EGW distance distribution is constructed, and the first predicted time is accurately identified through adaptive parameter optimization. Then, the ACIDAN is designed. Intradomain adversarial training is employed to eliminate feature-distribution shifts under multiple working conditions. The outputs of multisource regressors are coordinated through an interdomain kernel-density-estimation weighting strategy to ensure consistency. A multisource predictive variance method for the residual connected regressor is introduced in the degraded perception stage to generate confidence intervals containing uncertainty quantification. Finally, the RUL prediction accuracy and uncertainty quantification of the algorithm proposed in this article are verified.