Research On Flow Field Reconstruction And Suction Control Of Supersonic Diffuser Cascade

By reasonably organizing the shock wave structure in the cascade channel and increasing the single stage pressure ratio of the cascade,the supersonic cascade can reduce the number of stages and parts of the compressor and improve the working stability of each part of the compressor,and then significantly improve the thrust-weight ratio of the aeroengine.However,the multiple reflection of the shock wave in the cascade channel will lead to the thickening of the boundary layer,enhance the shock-boundary layer interaction and induce the flow separation phenomenon,which can easily lead to the deterioration of the flow field and even the surge phenomenon when the back pressure is high,which shows that it is of great practical significance to monitor the flow field parameters in supersonic cascades.Considering that the supersonic cascade flow field has strong distributed parameter characteristics,the traditional flow field state monitoring methods mostly use the information of a single pressure measuring point,which is difficult to directly and comprehensively reflect the status of the cascade flow field.Considering the strong nonlinear feature recognition ability of deep learning,and its outstanding achievements in the fields of pattern recognition,image analysis,decision judgment and so on.Therefore,this thesis focuses on the identification accuracy of datadriven model for supersonic cascade flow field and its influence on active flow control decision-making.The ability of the data-driven model to reconstruct the high spatial resolution flow field according to the wall discrete pressure is analyzed,and the enhancement effect of the data assimilation method on the time series prediction model is further studied.Then the hardware deployment scheme of the data-driven model is established.The real-time performance of the flow field reconstruction is verified in the ground wind tunnel experiment.Finally,the on-line cascade flow field reconstruction scheme is deployed in the suction active flow control experiment.In order to explore the ability of deep learning model to identify cascade flow field,the technical route of datadriven model hardware deployment and its help to flow control decision-making,the following aspects of work have been carried out:Firstly,the intelligent reconstruction of supersonic cascade flow field is carried out to estimate more abundant flow field from limited parameter measurements.The intelligent reconstruction framework of supersonic cascade flow field is established.According to different working conditions,the fusion model including transpose network and residual network,and the three-dimensional spatial sequence reconstruction model Conv LSTM are constructed respectively.The functional relationship between the discrete pressure point on the cascade channel wall and the cascade flow field,and the functional relationship between upstream pressure and downstream pressure in the cascade threedimensional flow field are established respectively.The original data for the training model are obtained by performing the free jet experiment and numerical simulation of the supersonic cascade.Through the analysis of the reconstructed flow field,we can know that the data-driven model based on deep learning can accurately identify the influence of the incoming flow condition and back pressure on the shock wave structure of the flow field,and accurately reconstruct the local pressure downstream according to the pressure of the upstream local region.Secondly,the flow field parameter prediction and model enhancement of supersonic cascade are carried out.In view of the complex spatial-temporal effect of cascade flow field prediction,a time sequence reconstruction model of embedding convolution operation into LSTM computing nodes is established.Free jet experiments with different flap movement rates are carried out to obtain the original data set.Through the analysis of pressure-schlieren and pressure-pressure sequence prediction results,it can be known that,the time sequence reconstruction model can accurately predict the flow field structure and wall pressure in the future finite time period according to the flow field parameters in the past finite time period.Furthermore,to improve the problem that the data driven model under the supervised learning paradigm is difficult to accurately capture the flow characteristics outside the training set,a Bayesian data assimilation framework is established.The ensemble Kalman filter is used to update the bias parameters of the original time sequence reconstruction model online,an updated flow field parameter prediction model is obtained,and the prediction results of wall pressure for future time steps are improved.Thirdly,the real-time on-line reconstruction and demonstration verification of supersonic cascade flow field are carried out.The data preprocessing methods and data cache queue mechanism in online reconstruction flow field experiments are introduced.The data-driven model of on-line reconstructed cascade flow field is generated.Through model quantization and compilation,the model is further deployed on the edge computing device Edge TPU.In order to transfer the wall pressure collected by the pressure sensor to the computing device in real time,the TCP/IP wireless data communication path is established,which ensures the real-time and accuracy of pressure data transmission.In the ground experiment,the accuracy of real-time flow field reconstruction under different working conditions is verified,the speed of flow field reconstruction by different hardware equipment is compared,and the real-time performance of flow field reconstruction is verified.Finally,application research on shock position estimation and control of supersonic cascade flow field was carried out.The real-time online flow field reconstruction framework proposed previously is deployed in the suction flow experiment to process the image of the real-time reconstructed flow field and extract the leading edge position of the shock train to judge whether the suction control instruction is initiated or not.The influence of suction on the flow field structure of supersonic cascade is studied.The layout of suction and the evolution of cascade flow field structure under suction control are introduced.Further,the influence of suction flow on the suction control effect is analyzed.The influence of suction active flow control on stability margin and maximum anti back pressure ratio is studied.The control effect is further compared with that of the traditional shock position estimation method(pressure ratio method).The results show that the shock position estimation flow control method based on deep learning can accurately judge the shock position of the flow field and turn on the suction control in time and suppress flow separation.In addition,the problem of control failure of pressure ratio method in the case of fewer sensing points and rapid rise of back pressure is improved.