A Fatigue Life Prediction Method Based on Optimized Data Augmentation Algorithm and Adaptive Sampling Model

ABSTRACT Machine learning–based fatigue life prediction is essential; yet their datasets typically contain fewer than 50 samples, severely challenging predictive modeling. This study proposes an integrated framework of simulated annealing (SA)‐guided data augmentation and residual‐driven adaptive sampling (RAS) bagging ensemble. SA optimization identifies augmentation parameters that maximize RAS‐bagging performance, while RAS adaptively reweights bootstrap sampling toward high‐residual regions to enhance ensemble diversity. Repeated nested cross‐validation (tenfold, 30 repetitions with identical splits) showed significant improvements over classical bagging on the CFRP () and DH36 () datasets (Wilcoxon signed‐rank test, all ; large effect sizes, Cliff's ). Mean improved from to and from to , respectively, while mean MAE was reduced by 85.8% and 64.3%, and the standard deviation of decreased by more than an order of magnitude (0.57–0.02 and 1.48–0.13). In addition, the 80% conformal prediction intervals derived from the proposed model achieved an empirical coverage rate of 84%, providing conservative lower‐bound life estimates for reliability‐oriented decision‐making. These results demonstrate that physically bounded data augmentation combined with adaptive ensemble resampling effectively alleviates data scarcity in fatigue life prediction within the studied datasets. Validation on larger independent cohorts, field data, and online learning scenarios is warranted before the framework can be extended to automotive reliability assessment and structural health monitoring.