Fatigue Life Prediction Method Integrating Adaptive Exponential Interpolation Data Augmentation and Machine Learning

Abstract

To address empirical parameter dependence, high test costs, and the accuracy‐applicability trade‐off in traditional fatigue cumulative damage models, this paper proposes a novel fatigue life prediction method integrating adaptive exponential interpolation data enhancement, an improved nonlinear damage model, and machine learning. Based on limited test data, virtual loads are constructed to boost stress‐domain sampling density at no extra cost. An improved AEVILD–Manson model with reliability criterion and convergence proof is established via Monte Carlo correction. Nine machine learning algorithms are integrated for adaptive calibration of load interaction coefficients. Verified by 12 material datasets and B750L steel multilevel fatigue tests, the model achieves absolute error less than 2.5% at 85% reliability; the R 2 value of the LSTM integrated model is 0.9668, and the RMSE is 0.0536, outperforming traditional benchmarks. This method enables high‐precision small‐sample fatigue assessment for low‐cost, high‐reliability structural design.