A dual-enhanced random forest model for fatigue life prediction under multi-level variable amplitude loading

To address two limitations (hyperparameter sensitivity and black-box fitting) of random forest (RF) models in predicting fatigue life under multi-level variable amplitude loading, a dual-enhanced random forest model (Dual-ERF) is proposed. First, Q-Learning reinforcement learning is introduced to improve the Langevin Equation Evolution (LEE) model, forming the Q-LEE model. Validated against benchmark test functions, Q-LEE model exhibits superior fast convergence and optimal stability compared to LEE model, enabling precise tuning of key hyperparameters for the random forest. Second, referencing the modeling framework of intrinsic damage dissipation, a new fatigue cumulative damage model incorporating nonlinear terms is proposed. Validated against material-level and structure-level tests, the new model has good performance compared to the classical models in prediction accuracy and provides a theoretical basis for establishing high-information-density input features. Based on above two improvements, Dual-ERF model shows excellent prediction accuracy and performs robustly in small-sample scenarios, when compared with test values. This study provides a highly accurate model for fatigue life prediction of metallic materials under multi-level variable amplitude loading.