Utilizing convolutional neural networks for calculating full-field stress components and directions in photoelasticity

Abstract

Photoelasticity is a crucial experimental technique extensively used in various engineering and scientific domains. The integration of convolutional neural networks can substantially enhance the efficiency and capability of photoelasticity in resolving full-field stress. Nevertheless, existing neural network–based methods in photoelasticity are limited to computing only the difference between the principal stresses rather than determining the individual principal stress components (i.e., the absolute values of the first and second principal stresses) and the principal stress direction (isoclinic parameter). Directly solving for the principal stress components is of greater significance in many practical problems, and the principal stress direction is essential for evaluating material failure and optimizing designs. In this paper, a convolutional neural network is proposed for the first time that can directly and simultaneously determine the full-field first principal stress, second principal stress, and the principal stress direction. A dataset generation method was developed to train this network, producing a novel high-quality dataset containing 41,000 raw samples, without data augmentation. The proposed network exhibits high accuracy and strong generalization across synthetic and experimental validation sets. On the synthesized dataset, the structural similarity exceeds 0.98, and the mean squared error is below 0.45, with similarly satisfactory results on the experimental validation sets. This network establishes a direct mapping between photoelastic images and full-field stress components and directions, thereby enhancing the efficiency and potential applications of photoelasticity. The proposed dataset generation method may also offer valuable insights for advancing deep learning in photoelasticity.