A data-driven multi-flaw detection strategy based on deep learning and boundary element method

In this article, we propose a data-driven multi-flaw detection strategy based on deep learning and the boundary element method (BEM). In the training phase, BEM is implemented to generate the database, while the block LU decomposition technique is employed to reduce the computational cost. Then the Convolutional Neural Networks (CNNs) are adopted as a deep learning model to find the relationship between the input signals and the geometries of flaws through the training process. In the test phase, the performance of trained models will be evaluated with unseen data. As a typical inverse problem, the solution to a flaw detection problem is not always unique. In the present work, we demonstrate that such non-uniqueness is detrimental to the training process, and avoid them through some specific treatments. In order to enhance the robustness of the model, the idea of data augmentation is introduced to flaw detection tasks. The numerical results show that the presented model could produce accurate predictions in both single- and multi-flaw detection tasks with proper training. Additionally, data augmentation could significantly help against the noise.