DEEP LEARNING APPROACH TO EVALUATE FRACTURE PARAMETERS FROM PHOTOELASTIC IMAGES

Since the inception of Fracture mechanics, the ability of photoelastic techniques to show significant geometric changes in the isochromatic fringe field has directed the mathematical modeling of stress field near the crack tip. The evaluation of fracture parameters, viz., stress intensity factors (SIFs) and T-stress is of utmost importance in predicting crack growth directions and estimating the life of the component. The current state-of-the-art technique for fracture parameters' evaluation uses the photoelastic fringe data to evaluate the coefficients of a multi-parameter stress field equation by minimizing the convergence error iteratively in a non-linear least squares sense. This is a multi-stage, semi-automatic approach. In this paper, the power of convolutional neural networks (CNN) that are well suited for recognizing complex spatial patterns is exploited to fully automate the evaluation of fracture parameters with the isochromatic image as the input. The network is pretrained on a large volume of simulated dataset which can later be fine-tuned for smaller experimental dataset. This approach helps to circumvent the requirement for large experimentally labeled dataset which is hard to obtain.