Inversion and reconstruction of supersonic cascade passage flow field based on a model comprising transposed network and residual network

Abstract

A data-driven model comprising a transposed network and a residual network is proposed to predict the flow field structure of the supersonic cascade passage by measuring the wall pressure of the cascade passage. This model is based on the close relationship between the wall pressure values of the supersonic cascade passage and the flow field structure. Wind tunnel experiments of the supersonic cascade passage were conducted to obtain real experimental data. The transposed network, which is the first part of the model, consists of transposed convolution, normal convolution, and fully connected layers. However, the mapping ability of the transposed network between the wall pressure values and the cascade passage flow field structures was insufficient. The maximum relative error by the transposed network in predicting the position of the leading edge of the shock train was 8.5%. Therefore, the residual network was introduced as the second part of the model to achieve end-to-end super-resolution and improve the mapping ability of the model. After training, the residual network reduced the maximum relative error of the position prediction of the leading edge of the shock train to 3.5%. In the testing set, the prediction results of the model indicated that the movement of the leading edge of the shock train can be distinguished well. Furthermore, a combination of the two networks revealed the mapping relationship between the wall pressure of the supersonic cascade passage and the flow field structure. Thus, the results of this study lay a good foundation for the subsequent determination of the flow field stability margin.