Radiation Source Recognition Based On Deep Learning And Implented By FPGA

With the development of communication technology,Individual radiation source identification(IRSI)is a common technique in many fields,such as communication countermeasures,wireless spectrum management,life sciences,and fault diagnosis.The challenges of the IRSI mainly come from the diversity of the radiation source,because of its multiple modulation methods,different center frequencies,and various transmission rates.Due to the little difference between radiation sources,the traditional method of IRSI based on machine learning cannot realize the requirement of accuracy,proposed by modern practical engineering.And the same is its complex structure and long recognition time.To solve the above problems,the deep convolutional neural network(DCNN)is applied in this thesis.Besides,a hardware accelerator characterized by high accuracy and low power is designed to meet the requirement of the high computational complexity and large amount of calculation of DCNN.Subsequently,this article describes the theoretical knowledge of deep learning,briefly summarizes the research background and development history of deep learning,and expounds the basic concepts of deep learning.And according to the classic structure of the convolutional neural network,the convolutional layer,pooling layer,fully connected layer and activation function of the convolutional neural network are analyzed in detail.Next,the thesis describes the generation mechanism of radiation source fingerprints,the common methods of radiation source feature extraction and traditional machine learning classification algorithms,and elaborates the radiation source recognition algorithm used in this article,which combines short-time Fourier transform and deep convolutional neural network.The traditional recognition algorithm and the algorithm of this thesis are used to identify the measured radiation source data sets.The comparative analysis proves that the network model of this thesis has better recognition efficiency in the recognition of individual radiation sources.In order to meet the low latency and low power consumption requirements of radiation source identification in practical engineering,this thesis designs a hardware accelerator for the radiation source identification algorithm.Describes the overall architecture of the radiation source individual identification system and the data interaction process of the system.It also describes the hardware architecture and work flow of the neural network processor core in the recognition system,and introduces the neural network processor in this article in detail from the main control unit,interactive unit,storage unit,computing unit,and shaping unit.And it is implemented on a hardware platform with ARM+FPGA structure.Finally,for the individual radiation source identification system constructed in this article,the measured radiation source data set is used for experimental evaluation on the ZYNQ-7045 development board.The algorithm identification accuracy rate is 96.50%,the hardware identification accuracy rate is 95.57%,and its hardware implementation accuracy The loss is only 0.93%.And the measured board-level power consumption is2.746 times the lowest energy efficiency ratio compared with other commonly used GPUs.