Predictive Fatigue Analysis Using Synthetic Data and Deep Learning Models

Abstract

This thesis explores a hybrid methodology combining finite element simulation and machine learning techniques to predict fatigue crack growth in metallic structures. Inspired by the work presented in *A Framework for Generating Large Data Sets for Fatigue Damage Prognostic Problems* (FDPP), this research investigates the credibility and effectiveness of synthetic data by comparing it with results obtained from high-fidelity simulations in COMSOL Multiphysics. A 2D fatigue crack model was simulated under cyclic loading, with strain gauges placed in accordance with the FDPP setup. The resulting strain and displacement data were processed using custom Python scripts to derive meaningful comparisons between the synthetic and simulated datasets. The primary contributions include validating the synthetic dataset’s realism via numerical comparison, and introducing alternative machine learning models—such as XGBoost, LightGBM, and Autoencoders—for Remaining Useful Life (RUL) estimation. The findings demonstrate a promising correlation between real and synthetic data trends, paving the way for more credible fatigue prognostics using virtual datasets in conjunction with physics-based simulations.