Deep Learning Intelligent Signal Processing In Radar Based On FPGA

The radar intelligent signal processing technology can not only simplify the operation but also more effective operating results.In this thesis,we take rotorcrafts as the targets of recognition to study the intelligent target recognition technology based on deep learning.Rotorcraft play an important role in both military and civilian fields.By extracting and analyzing the characteristic information in the radar echo signal of the Rotorcraft,we can effectively recognize types of Rotorcraft.The rotation of the Rotorcraft rotor will cause micro-Doppler,and the micro-Doppler feature can be obtained from the radar echo using joint time-frequency analysis.Convolutional Neural Network(CNN)has excellent performance in feature extraction.Using CNN can extract feature information from the micro-Doppler feature map and accomplish tasks of Rotorcraft recognition.However,with the increase in network scale,it is difficult for the central processing unit(CPU)to bear the huge amount of calculations.It is very necessary that accelerating CNN on hardware.Field programmable gate array(FPGA)has the advantages of high parallelism,low power consumption,and high flexibility and is very suitable for accelerating CNN.According to the above background,this thesis uses CNN to recognize the category of rotorcraft basing on micro-Doppler,and accelerates the processing on the FPGA to achieve the rotorcraft recognition task more efficiently.The main contents and contributions are as follows:(1)Modeling and analyzing the micro-motion characteristics of rotorcraft,deducing the radar echo signal model of the horizontal and vertical rotors,and analyzing the effect of the actual characteristic blade angle of the rotor on the echo.On the basis of the micro-motion model,simulation experiments are carried out on helicopters,jets and quadrotor.Taking advantage of the short-time Fourier transform in time-frequency filed,the measured echo data is processed to obtain the micro-Doppler feature maps of helicopters,large quadrotor,and small quadrotor.The analysis of dynamic characteristics lays the foundation for the recognition of rotorcraft in the following text.(2)Using CNN to extract the micro-Doppler features of the aircraft,and achieving the purpose of aircraft recognition based on the feature difference.CNN can effectively obtain the feature information in the micro-Doppler feature map to classify and recognize rotorcraft.In order to improve network performance and accelerate implementation on the FPGA,small convolution kernels replace large convolution kernels,and dilated convolutions replace part of ordinary convolutions,and global pooling is added before fully connected layers.The above three aspects are used to optimize the network.The number of optimized network parameters is significantly reduced,the amount of calculation of convolution point multiplication is reduced by nearly half,and the accuracy of network recognition is improved,which proves that the optimization scheme is effective and feasible.(3)According to the structure characteristics of the optimized network,the acceleration scheme on FPGA of optimized network is designed.It is designed how implete each submodule on hardware.The convolution part includes ordinary convolution and dilated convolution.Pooling includes maximum pooling and global average pooling.Deign the storage of the initial input feature map,network parameters and intermediate results.The convolution part of the data to be calculated is buffered using a shift register.Each convolution operation only needs to read in data that is different from the previous clock,which greatly reduces the time consumed for repeated reading of data.The input data and intermediate results are separately stored in DDR and FIFO.The comparison of time-consuming,resource utilization,power consumption and other aspects with the running results of the CPU platform verifies the effectiveness of the FPGA-based acceleration scheme in this paper.