Radar Clutter Recognition Based On Convolutional Neural Network Design And Hardware Implementation

Constant False Alarm(CFAR)is one of the key technologies in the radar signal processing process,which can maintain a constant false alarm rate during target detection.Traditional constant false alarm processing is known as a precondition for the statistical distribution characteristics of clutter.A constant false alarm processing method is usually only applicable to a clutter environment.With the development of sea,land and air integration,the working area and environment of the radar often change.The traditional constant false alarm processing method is difficult to effectively deal with the clutter of various statistical distribution characteristics.If the constant false alarm detector can identify different radar clutter and make corresponding processing according to different clutter environments,the constant false alarm detector can be adaptive to the multi-clutter background.Therefore,the research on radar clutter classification and recognition method has profound theoretical significance and application value.In this thesis,it has been proposed that a method based on convolutional neural network radar clutter recognition by studying convolutional neural networks in the field of image processing.Generally,Rayleigh clutter,Weibull clutter,lognormal clutter and K-distribution clutter are common radar clutter in radar system processing.Four types of radar clutter are generated by using ZMNL clutter simulation to form a clutter data set for convolutional neural network training and testing.In the algorithm modeling stage,by selecting six network structures with different convolutional layers,under the same training test data set,according to the training test accuracy of the six networks,the number of operations of the convolution layer,the parameters of the convolution layer A comparative analysis of the quantities determines the network structure used in this paper: the network structure of the four convolutional layers.After training and testing,the classification accuracy of the four types of radar clutter used by the network used in this paper is 99.51%.Using Deep Dream technology to visually analyze the trained convolutional neural network,the feasibility of applying convolutional neural network to radar clutter identification is demonstrated.After the algorithm modeling is completed,the convolutional neural network of radar clutter recognition is implemented based on FPGA platform.In this paper,the hardware and software collaborative design flow is used to implement the hardware implementation of the algorithm,including the design,verification,selection of hardware optimization scheme,functional verification of RTL algorithm model,logic synthesis and board level verification after place and route.In the hardware implementation stage,four different hardware optimization schemes are selected by analyzing the data flow and operation mode of the convolutional neural network.The performance and resource occupancy of different hardware optimization schemes are determined to determine the hardware optimization scheme adopted in this paper.In the functional verification phase of the RTL algorithm model,the C/RTL collaborative verification method is adopted to ensure that the RTL algorithm model and the C algorithm model are consistent.The MATLAB algorithm model has the same recognition accuracy under the same test data set.After logic synthesis and place and route,the hardwareimplemented algorithm model is verified on the ZC706 platform.The verification results show that under the same test data set,the hardware implementation algorithm model is consistent with the recognition accuracy of MATLAB algorithm model.It shows that the design can classify and recognize the radar clutter background,showing good adaptability and robustness.