Deep learning based segmentation of ice pore structure in icing wind tunnel experiments

Ice formation poses a serious threat to the flight safety of aircraft, and accurately identifying the three-dimensional pore structure in aircraft icing is of great significance for icing simulation and de icing design. Traditional segmentation methods based on convolutional neural networks often perform poorly at block boundaries, resulting in discontinuous and incomplete pore recognition. Therefore, this article proposes an automatic 3D segmentation method based on deep learning architecture, aiming to improve both local segmentation accuracy and global structural continuity. This method is based on the 3D U-Net architecture, introducing a Mamba based feature extraction module to optimize intra block segmentation performance, and introducing a Transfromer based feature extraction module to maintain the connectivity of pores at block boundaries. The pore structure segmentation of typical frost ice and clear ice was carried out using ice obtained from small-scale icing wind tunnel tests. The study shows that: (1) the method achieved 92% F1 score and 98% accuracy, significantly better than the 89% and 94% of the standard 3D U-Net. (2) Clear ice has few dense pores that are not interconnected, while frost ice has abundant pores with strong connectivity and high permeability.