Deep reconstruction of schlieren images for scramjet combustion flow field analysis using FCSD-net with feature completion strategy

Abstract

Reconstructing schlieren images of flow fields using wall pressure measurements is widely regarded as an effective technique for monitoring the state of flow fields in scramjet engines. However, this reconstruction process can be significantly affected by wave system changes resulting from flow field oscillations and boundary layer interference. To tackle this issue, we introduce a dual-branch network structure based on a feature completion strategy, known as FCSD-Net. FCSD-Net integrates a transformer-based primary branch with a convolutional neural network-based auxiliary branch to capture multi-dimensional semantic features. Additionally, a fourth-order feature fusion module is designed to effectively combine global features. Experimental results indicate that FCSD-Net can efficiently reconstruct schlieren images. This method achieves improvements of 5.54 % in SSIM, 3.02 dB in PSNR, and 2.4 % in P-corr compared to leading neural network models. The proposed approach accurately reconstructs wave system structures and turbulence in supersonic flow fields, providing a solid data foundation for analyzing flow field stability in combustion.