Intelligent reconstruction of supersonic combustion flow fields using sparse sensor data and heuristic-guided learning

Reconstructing the global combustion flow state under supersonic combustion conditions using sparse sensor data presents a novel and significant challenge. This challenge is critical for advancing measurement and diagnostic technologies in supersonic vehicles operating in complex and extreme environments. However, the task becomes notably tricky when the number of sensors is minimal. Current end-to-end learning models often face significant generalization issues when reconstructing combustion flow across the entire spatial domain, especially in the context of sparse pressure measurement systems typically used in ground wind tunnel tests. To this end, this study introduces an intelligent reconstruction model for combustion flow that incorporates heuristic-guided learning, enabling accurate flow field reconstruction across the entire spatial domain using sparse pressure sensor measurements. The proposed model operates in two distinct stages. The first phase is a heuristic learning phase, which uses the multi-gradient learning strategy to learn the joint feature distribution of different scales of the flow field in the whole spatial domain based on the sparse pressure measurement data. In the second stage, the guided learning stage, a multi-scale feature fusion mechanism is applied to refine both the content and structural details of the coarse and fine distribution features within the joint feature distribution. The efficacy of the proposed model is validated using a ground test dataset collected at various Mach numbers. Experimental results demonstrate that the model achieves superior performance across various complex and extreme scenarios.