Deep Learning for Image Segmentation and Subsurface Damage Detection Based on Full-Field Surface Strains

Damage detection plays a key role in estimating the health of the structure. Accurate damage detection allows judging the reduced capacity of the structure and further retrofitting the damage. Error in damage detection may have catastrophic consequences, especially when the damage is subsurface that may accumulate over time. In this study, a deep convolutional neural network (CNN) based on full-field strain measurements is developed to localize subsurface damage. The dataset is prepared artificially by finite element simulation of rectangular metal bars having subsurface damage of varied length, size, depth, and direction of propagation. For the trained network, the Intersection of Union score is found to be 0.72 for both training and testing set. This implies that the model can localize the subsurface damage and can be further explored for applications in nondestructive testing. For continuously generated strain maps, applications in dynamics systems to study damage initiation and propagation can be studied for dynamic loading.