Research On Digital Signal Modulation Recognition Technology Based On Deep Learning

With the rapid development of Internet of Things(IoT),the communication of machine type are increasing.Moreover,the integration of mobile services and machine-type communications ushered in the era of wireless communications big data.However,this has also led to an increasing number of various low-power signals in the spectrum environment,which intensifies the complexity of wireless electromagnetic environment.Meanwhile,the seamless IoT unveils a number of physical-layer threats,such as jamming and spoofing.The effective digital signal modulation recognition technology can provide higher-precision sensing and analysis capabilities for spectrum monitoring and management,and can detect and identify physical layer threats such as pilot interference and deceptive interference.Therefore,this thesis aims to use deep learning technology to research the recognition of digital signal modulation methods to improve the accuracy and robustness of digital signal modulation recognition.The main content of this thesis is as follows:1.Modulation recognition based on cyclic correntropy and long short-term memory densely connected networkIn this thesis,we propose a signal modulation recognition method based on cyclic correntropy vector(CCV)and long short-term memory densely connected network(LSMD).This method uses deep learning technology and cyclic correntropy vector with strong characterization ability as the original feature,and it can be divided into two stages,the offline training stage and the online recognition stage.In the offline training stage,we adopt a large amount of CCV feature data to drive the training of the LSMD network and get the final LSMD classifier.In the online classification stage,we first calculate and extract the CCV feature of the signal obtained from the receiver,and then input the CCV features into the LSMD classifier,and finally output the decision result.In the LSMD network design,we use the additive cosine loss function to maximize the inter-class difference and minimize the intra-class difference,so as to improve the classification and recognition ability of the LSMD classifier.The simulation results verify that this method can achieve a performance gain of 3.8dB when the recognition accuracy rate is 95%compared with the optimal comparison scheme.2.Modulation recognition method based on multi-module fusion neural networkIn order to solve the problem of irreversible information loss caused by the manually designed statistical features,a signal modulation recognition method based on multi-module fusion neural network is proposed.This method uses pixel-coloring constellation projection algorithm(PCCP)with very low computational complexity to obtain enhanced pixel-coloring constellation features.Then,we designed a multi-module fusion neural network(MMFN),which includes three modules of feature conversion,representation and classification,to automatically extract high-dimensional representations of different modulated modes through feature conversion and representation modules.In addition,the learning ability of the MMFN network is greatly enhanced by adopting the design idea of multiple cross-layer combinations.A large number of simulation experiment results prove that when the signal-to-noise ratio is 0dB,the recognition accuracy of this method is improved by 7%compared with the optimal comparison scheme.