Research On Intelligent Modulation Recognition Of Communication Signals And FPGA Implementation

In recent years,with the rapid development of wireless communication technology,the modulated signals in wireless communication systems have become increasingly complex and diverse.Among the many parameters of modulated signals,the signal modulation method is known as a prerequisite for signal demodulation at the receiving end.However,in the non-cooperative communication scenario,the receiver cannot be informed of the modulation mode of the communication signal in advance due to the non-cooperative nature between the transmitter and the receiver.Although studies have been conducted to achieve modulation recognition of signals based on methods such as higher-order statistics of received signals,the performance is limited by the reasonableness of feature and threshold selection.Machine learning-based modulation recognition can automatically extract signal features without complicated threshold selection,and therefore has gained industry attention in recent years.In this thesis,modulation recognition of communication signals and FPGA implementation are studied,and the specific work and contributions are as follows.First,this thesis studies the traditional modulation recognition method based on feature extraction.The extracted features mainly include timefrequency domain features related to the spectral density,high-order cumulative features and constellation diagram features.Based on the above features and the selection of suitable thresholds,this thesis realizes the identification of various modulation methods including FSK,PSK and QAM modulations.The simulation results show that the method has excellent recognition performance at high signal-to-noise ratio.However,this method relies heavily on threshold selection,and if the threshold is not selected properly,the performance will drop abruptly.Secondly,this thesis investigates the modulation recognition scheme based on convolutional neural network and long-short time neural network,which effectively identifies the unknown modulation mode of the signal;however,the parameter number of this network model is too large and difficult to deploy in practice.In order to compress the network model,this thesis proposes a lightweight neural model combining random quantization training and iterative product quantization.Simulation results show that the proposed scheme can achieve 17.5 times the weight compression ratio compared with the uncompressed neural network model while ensuring the recognition accuracy.Finally,two modulated recognition neural networks,i.e.,compressed and uncompressed,are implemented on an Artix7 FPGA chip in this thesis.Simulation experiments show that compared with the uncompressed model,the compressed model has a 37.1%reduction in storage resource BRAM usage.Moreover,the recognition speed of the compressed model is 5.3 times faster than that implemented on the CPU.In summary,the neural network implemented in this thesis does not require complex feature and threshold selection,the proposed compressed model can significantly reduce the resources required for network storage,and the FPGA deployment platform can effectively improve the recognition speed of the modulation method.Therefore,the modulation recognition scheme studied in this thesis has great practical application value and contributes to noncooperative communication.