Research On Measurement And Monitoring Method Of Scramjet Based On FPGA And Deep Learning

Because of the important strategic position,hypersonic propulsion technology has been widely studied in many countries including China.With the technology becoming more and more mature,it puts forward practical requirements for the intelligent management system of scramjet,which is embodied in the increasing requirements for the measurement and condition monitoring of key parameters.The method based on deep learning can effectively solve those problems.But deep learning neetwork has the characteristics of large amount of parameters and calculation,which makes it difficult for traditional hardware platform to carry out the algorithm.Based on the above problems,this paper proposes a hardware platform of scramjet based on Field Programmable Gate Array,FPGA.The main contents are as follows:In this paper,zynq heterogeneous SOC is selected as the hardware platform of convolutional neural network deployment according to the constraints in the application of hardware platform of scramjet and the structural characteristics of convolutional neural network model.Based on the concept of hardware software co design,the overall scheme of hardware acceleration system of convolutional neural network is proposed.Under the guidance of the overall scheme,the design of the whole CNN accelerator system is completed.Firstly,the task division and overall architecture design of CNN accelerator system based on FPGA are carried out in combination with CNN network model.In this paper,we design several functional sub modules of CNN accelerator,make full use of FPGA on-chip resources,and give full play to the parallel ability of convolutional neural network operation.In view of the circulation in convolution layer and pooling layer,block and expansion strategies are adopted,and the block parameters are selected based on the theory of rootline model.In the further optimization,the different hardware optimization technology and the feasibility under the design requirements are studied,and the design of the PL programmable logic part is completed.At last,the data transmission of DMA module and the software development of PS terminal are designed.The performance of CNN accelerator is verified,and the convolutional neural network is used to test the performance.The related verification of CNN accelerator is carried out on the development board of ZYNQ chip equipped with XCZU3CG-1SFVC784.The test results show that the resource consumption of CNN accelerator is reasonable,the overall power consumption is low,and it has good acceleration performance.Compared with CPU platform and other designs,the advantages and disadvantages of the accelerator are summarized.Further,aiming at the specific application of CNN accelerator,two CNN models are verified.The test results show that the operation results of CNN accelerator have good accuracy,can reflect the actual flow field,the distribution of parameter field,and has a certain degree of reality,to meet the use needs.The intelligent measurement and control system of scramjet is studied.Based on the measurement and monitoring system framework of scramjet equipped with ZYNQ and CPU processor,the front position detection system of shock train in the isolation section of scramjet is designed.The data transmission between CPU platform and ZYNQ platform is realized by using high-speed data transmission protocol.ZYNQ receives the measurement information to complete the deep learning algorithm.CPU platform processes the data and completes the Information output.According to the subsequent control requirements,the engine model and controller are installed on the ZYNQ development board,and based on this,the closed-loop control model on the development board is constructed.