Stress field identification using deep learning and three-dimensional digital image correlation

The stress borne by adhesive layers has drawn much attention to adhesive assembly since considerable deformations are introduced to bonded structures. Rapid analysis of stress and deformation plays an important role in adjusting the boundary conditions and optimizing the design of bonded structures. This paper develops a stress identification method to predict the stress distribution and evolution in adhesive layers by integrating a multitask stress learning-Unet (MTSL-Unet) model, a three-dimensional digital image correlation (3D-DIC) technique, and a finite element model. In this method, the MTSL-Unet model is trained using a dataset that includes the surface deformations acquired via 3D-DIC and the corresponding stress fields obtained via FEM simulations. Finally, an end-to-end prediction is achieved from the surface deformation to the stress distribution in the adhesive layers. For dynamic stress identification, sequential surface deformations of a bonded structure are obtained via a custom 3D-DIC system, and the stress distributions in the adhesive layer are accurately and rapidly determined via the MTSL-Unet model. The presented method is a technique that has potential for use in real-time noncontact stress monitoring and is suitable for digital twins in mechanical and structural engineering.