Abstract
Traditional FRP durability forecasting predominantly treats performance evolution as a discrete "point-to-point" regression, inherently overlooking temporal coherence and stochastic uncertainty. This study proposes a novel probabilistic framework based on the CSDI diffusion model to reconstruct continuous FRP degradation curves. By formulating long-term forecasting as a conditional imputation task, the methodology generates physically consistent performance trajectories from sparse experimental observations. Results from a multi-factor database demonstrate that CSDI enables a paradigm shift to continuous sequence generation, achieving high predictive accuracy (RMSE = 0.332, R(2) = 0.86) and robust probabilistic calibration (CRPS = 0.170) at a 30% missing ratio. This approach establishes a reliable probabilistic risk envelope, providing a scientific tool for the life-cycle reliability assessment of FRP structures under small-sample constraints.