Implementation Of Digital Signal Modulation Recognition With FPGA Based On Machine Learning

Signal modulation identification refers to making a correct judgment on the modulation type of the received signal when the modulation information is unknown.Modulation recognition of digital communication signals is not only of great significance in the civilian field,but also plays a vital role in military and national security.In electronic warfare,quickly and accurately determining the modulation of unknown signals is the key to victory.Under the wave of artificial intelligence,the combination of machine learning algorithms and modulation recognition technology has become a development trend.The modulation recognition algorithm based on machine learning and the hardware implementation method of neural network is studied in this thesis,in order to realize the recognition of digital modulation signals on the FPGA development board.The neural network built by hardware can not only almost achieve the recognition accuracy of the software-implemented network,but also the training speed is much higher than the rate of the software.The need for using intellectual property(IP)cores is eliminated for all functional modules except for multipliers and adders,which is easy to implement the transplantation of different FPGA platforms.Firstly,a modulation recognition algorithm based on instantaneous features is studied,and a decision tree and a neural network classifier are used to classify the modulation signals.In the process of using the decision tree to classify the modulated signal,the threshold value of each characteristic parameter is determined through simulation and the signal is classified according to each parameter.The experiment shows that the signal recognition rate is higher under high signal-to-noise ratio.When the noise ratio is greater than 14 d B,the overall recognition rate is above 95%.The results of the modulation recognition algorithm based on the neural network show that the neural network has the advantages of high flexibility,high robustness,and excellent performance in the process of modulation recognition.Then,for some occasions with high real-time requirements and when the neural network processes a large number of samples or the number of layers,and the number of neurons in each layer increases,the training time cost will increase.The implementation of FPGA of the BP neural network based on the BP neural network algorithm is proposed in this thesis,BP neural network is implemented with hardware description language,the key problems in the hardware implementation process are analyzed and solutions are given,and the implementation process of each module is introduced in detail.Through the design of the state machine,each module can run in an orderly manner,and the correctness of the functions of each module is verified through simulation.Finally,the recognition of the digital modulation method is realized on the FPGA.Through the overall function simulation and board experiment,the hardware function is correct within the error tolerance.The experimental results show that under the same network structure and the hardware clock frequency is 50 MHz,the difference between the FPGA-implemented network and the software-implemented network's recognition accuracy of modulated signals is about 1%,and hardware training speed has increased by 2 orders of magnitude compared with software,which shows that the FPGA-based neural network computing architecture has important practical significance,and provides a new idea for occasions with high real-time requirements for signal modulation recognition.