Machine learning based thermal imaging damage detection in glass-epoxy composite materials

Machine learning (ML) based fatigue damage detection from thermal imaging in glass-epoxy composites is an important component of remote structural health monitoring used for safety assessment and optimization of composite structures and components. However, accurate characterization of fatigue damage hotspots in terms of size, shape, location, hysteretic heat, and local temperature deep inside the material using surface thermal images remains a challenge to date. This work aims at evaluating the theoretical accuracy level of hotspot characterization by training a ML model with artificially generated thermal images from 3D finite element models with increasing complexity. Modelling the fatigue damage as an intrinsic heat source allowed to significantly reduce the influence of thermal image noise and other uncertainties related to heat transfer. It is shown that ML can indeed accurately recover the heat influx, depth, and geometry of the heat source from the original thermal images of the composite materials with prediction accuracies in the range 85%–99%. The effect of training set size and image resolution on the prediction error is also presented. The findings reported in this work contribute to the advancement of accurate and efficient remote fatigue damage detection methods for fibre composite materials. •ML accurately predicts parameters of subsurface damage hotspots of fibre composites.•Prediction accuracies for hotspot parameters are in the range 85%–99%.•Combining ML with thermal models improves fatigue damage detection.