The remaining useful life (RUL) prediction of rolling bearings under frequent start-stop cycle conditions is affected by the frequent switching of working conditions, which results in the weakened correlations between features and degradation states as well as periodic fluctuations of degradation trajectory. These issues make it difficult for traditional methods to construct effective health indicators. Accordingly a virtual health indicator construction method based on SA-PSO-STL is proposed to address this issue. For the weakened correlation problem, a simulated annealing particle swarm optimization (SA-PSO) algorithm is used to screen sensitive degradation features strongly correlated with degradation states. This process extracts feature information highly relevant to the bearing degradation process, thereby enhancing the monotonicity, trendability, and robustness of degradation representations. Then the principal component analysis (PCA) is utilized to fuse the selected features. Furthermore in order to mitigate the interference of periodic fluctuations in the degradation trajectory, the seasonal and trend decomposition based on the loess (STL) is employed for extracting the trend and separating periodic fluctuations as well as residual noise, thereby obtaining the virtual health indicators that can clearly and stably characterize the entire life cycle degradation process. Finally an adaptive multi-stage degradation model based on the sparse gaussian process regression (SGPR) is performed for the remaining useful life prediction. The experimental validation is conducted using three sets of full life cycle bearing data under different frequent start-stop cycle conditions. The results show that the constructed virtual health indicator (VHI) achieves comprehensive evaluation scores of 0.955 5, 0.962 2, and 0.989 2 on datasets A, B, and C respectively, while the corresponding RUL predictions demonstrate root mean square errors (RMSE) of 0.106, 0.047, and 0.014, as well as mean absolute errors (MAE) of 0.012, 0.002, and 0.002, respectively. It′s found that all quantitative indicators are significantly superior to all comparison models, which demonstrate the effectiveness and innovation of proposed model in enhancing the characterization of degradation trends and improving the prediction accuracy of bearing remaining useful life under frequent start-stop cycle conditions.