Fatigue damage reconstruction in glass/epoxy composites via thermal analysis and machine learning: A theoretical study

This study introduces an advanced, non-contact diagnostic tool for structural health monitoring of fatigue damage in fiber/polymer composite materials. The approach combines thermal image recognition of fatigue self-heating hotspots with high-fidelity thermal modeling to quantitatively assess subsurface fatigue damage distributions by machine learning. To this end, artificial thermal images are generated through 3D numerical thermal analysis of an inherent fatigue damage heat source within a glass/epoxy composite, derived from sampling a multivariate Gaussian distribution of microcracks. Subsequently, these synthetic thermal images are employed to train three distinct regression models: a convolutional neural network, a Gaussian processes regressor, and a straightforward least squares model. Various image augmentation techniques are applied to expand the dataset efficiently. All models accurately predict the size of the damage and – most importantly – the maximum temperature within the damage deep inside the composite. The regression methods estimate the diagonal elements of covariance matrix components of the Gaussian distribution, with accuracies ranging from 86% to 99%. The findings presented in this work contribute to establishing a solid foundation for non-destructive subsurface fatigue damage assessment in composite materials, with many practical applications in experimental composites fatigue research. •The models can accurately predict the maximum self-heating temperature in the damaged region inside fiber/polymer composite materials.•The results suggest that the full reconstruction of the subsurface fatigue damage distribution is feasible.•Some independent key covariance tensor components of a multivariate Gaussian damage distribution are predicted with accuracy in the range 93%–99%.