A multimodal attention fusion-based model for velocity field prediction of supersonic isolators

Monitoring the internal flow state of a supersonic isolator is crucial for maintaining the stable operation of hypersonic propulsion systems. However, in wind tunnel experiments, only local wall pressure signals can be collected by pressure sensors, which fail to perceive the full-field distribution of internal velocity field parameters; meanwhile, although full-field data can be provided by Computational Fluid Dynamics (CFD) simulations, they are plagued by high computational costs. To address these limitations and achieve high-precision velocity field prediction, a multimodal data fusion model integrating experimental measurement information and CFD-derived ground truth data is proposed in this study. An innovative dual-channel feature encoding module is constructed in the model: a pressure encoder based on Long Short-Term Memory (LSTM) is used to extract local temporal flow features collected by the wall sensor array; a schlieren encoder combining multi-scale convolution and spatial attention mechanism is employed to capture global spatial structural information of the flow field. A Transformer-based cross-attention mechanism is introduced in the model to establish dynamic correlation weights between the two modalities, enabling effective fusion of complementary temporal and spatial features and avoiding the defect of simple feature concatenation in traditional multimodal models. This study confirms the proposed model effectively fuses complementary information from multi-source heterogeneous data, offering a high-precision, cost-effective solution for supersonic isolator monitoring and performance evaluation.