Intelligent reconstruction algorithm of hydrogen-fueled scramjet combustor flow based on knowledge distillation model compression

Abstract

The real-time perception of the combustion flow of a scramjet can help to quickly evaluate the working state and provide a new opportunity for the intelligent reconstruction of the data-driven combustion flow field. However, the working process of supersonic combustors is extremely short, usually of the order of milliseconds, and the inference speed is slow due to the large number of parameters of traditional deep learning flow field reconstruction models within a very short time. Currently, knowledge distillation is known as an effective model compression method. In this paper, a neural network model based on symmetric structure cascade (NNSSC1) is proposed. By means of symmetric structure cascade, features of different levels are spliced and fused, which can effectively improve the reconstruction performance of the model. Secondly, to accelerate the application of the neural network model in the flight test, a student model with a simple structure is constructed. By distillation learning, the NNSSC model is used as the teacher model, combining the high performance of the teacher model and the high efficiency of the student model, and the combustion flow field is reconstructed with high speed and high precision based on the supersonic combustor wall pressure. In a direct connected supersonic pulse combustion wind tunnel with an inflowing Mach number of 2.5, the model was trained and tested on a dataset constructed in a hydrogen fuel scramjet experiment with different equivalent ratios. This method achieves high-precision and efficient reconstruction of complex flow fields in the combustor based on sparse wall pressure, providing more reliable data for accurately evaluating the flow and combustion status of the combustor of a scramjet.