Abstract: To address the complex challenge of time-varying reliability assessment in engineering structures subjected to dynamic loads and material degradation, this study proposes an efficient analysis method based on recurrent neural networks (RNN). The proposed approach functions as a data-driven surrogate model that learns from historical response data to achieve high-accuracy predictions at low computational cost. Unlike traditional analytical models that rely on explicit response functions, the RNN architecture automatically captures the nonlinear coupling between static random variables and time-dependent stochastic processes. A static variable embedding mechanism is designed within the network, enabling time-invariant variables to participate in modeling at every time step, thereby enhancing the model’s ability to represent system responses over the entire life cycle. Additionally, the Huber loss function is introduced to improve training robustness while maintaining predictive accuracy. Numerical experiments show that the proposed method can closely approximate the results of 10⁵ Monte Carlo simulations (MCS) using a minimal number of training samples. The proposed approach can not only accurately estimate the minimum of the limit state function, but also achieve global prediction of the system response curve within any specified time interval, expanding its applicability in practical engineering scenarios. Robustness tests under various perturbation settings show that the average prediction error consistently remains below 0.25%, demonstrating strong generalization and stability. Overall, the proposed method maintains high prediction accuracy while significantly reducing sample requirements, enabling efficient and reliable modeling of time-dependent reliability in complex systems.