Extended damage reconstruction in composites based on deep learning and time reversal algorithm

Composites can achieve higher strength, stiffness and durability than single materials, and their excellent mechanical properties make them widely used in aerospace, automotive, marine and construction. However, during the preparation or use of composite materials, damages such as voids and cracks may occur, which seriously affect their performance. These damages usually have continuous and irregular shapes in real life, so it is important to reconstruct the shapes of these damages in real time and accurately. In this paper, the effectiveness of time reversal algorithms and the U-Net algorithm in deep learning in solving this problem were investigated respectively based on electromagnetic inverse scattering imaging of composite materials, and based on this, the TR-Unet algorithm was proposed, which utilized the results of multiple source time inversions as inputs to the model to predict the actual relative permittivity distribution of the detected region, and then reconstructed the shape of the damage. The experimental results showed that the TR-Unet method showed significant improvement in both shape similarity and accuracy of contour details compared to the other two methods.