Fatigue life prediction and periodic operation and maintenance optimization of high-strength steel welded joints considering residual stress

Abstract

Existing fatigue assessment methods for high-strength steel welded joints commonly provide isolated life estimates and cannot directly convert residual-stress effects and prediction uncertainty into operation and maintenance decisions. To address this limitation, an integrated RS-SN-Stack-RDF-MDP framework is developed. A total of 864 experimental fatigue-life records are divided by original test series into training, validation, and held-out test sets, while 196 records from non-overlapping publications are used for external validation. Residual-stress-corrected S-N life and crack-growth descriptors are combined with out-of-fold predictions from SVR, RF, XGBoost, LightGBM, and MLP, and fused by a physics-constrained Ridge meta-learner. Predicted life is converted into remaining life, damage, and next-period failure probability for five-state average-cost maintenance optimization. RS-SN-Stack achieves an RMSE of 0.058, MAE of 0.044, MAPE of 3.1%, and R2 of 0.947 on the internal test set, while the corresponding external results are 0.084, 0.064, 4.4%, and 0.879. RDF-MDP opportunistic maintenance reduces the long-term average cost from 168.4×103 to 126.5×103, failure penalty cost from 39.4×103 to 20.8×103, and average downtime from 19.2 to 11.8 h. The framework establishes a continuous computational link from residual-stress-corrected fatigue prediction to maintenance action selection. Its validated applicability is currently limited to the investigated high-strength steel grades, joint geometries, and loading ranges.