Research On Deep Learning Based Modulation Recognition Under Complex Scenarios

At present,the use of deep learning（DL）to recognize modulation type of the received signal has received extensive attention.Most of the existing modulation recognition algorithms based on DL assume that the environment is simple and fixed,which hardly meets the needs of practical applications.This thesis considers three complex and varying scenarios,namely complex signal-to-noise ratio（SNR）scenario,complex channel scenario and complex signal scenario,and investigates the DL based modulation recognition techniques under these scenarios.Firstly,considering the complex and varying SNR values,two algorithms are proposed in this thesis,including the second and fourth-order moments（M₂M₄）estimation based algorithm and the multi-label DL based algorithm.The former trains several DL models at different SNR values in advance,and adopts an M₂M₄ estimator to estimate the SNR of the signal,according to which a proper trained model is selected for modulation recognition.The latter explores the multi-label method to train one model,by which both the modulation type and the SNR value can be inferred simultaneously.Experimental results show that two algorithms,especially the multi-label DL based algorithm,can effectively suppress the impacts of varying SNR.Then,considering the complex and varying channel types,three algorithms of the channel identification based algorithm,the single-label algorithm and the multi-label algorithm are proposed in this thesis.The first algorithm utilizes two cascaded convolutional neural networks（CNNs）to identify the channel type and the modulation type successively.The latter two algorithms employ a single CNN to eliminate the effect of varying channel and complete modulation recognition.Experimental results show that three algorithms are able to significantly improve the accuracy of modulation recognition under complex channel scenarios.Among them.the multi-label algorithm performs the best,and the single-label algorithm is superior to the channel identification based algorithm.Finally,considering the existence of interferences along with the primary signal to be recognized,this thesis adopts three classical CNNs to achieve modulation recognition under complex signal scenarios.The impacts of interferences on the primary signal are also analyzed.Experimental results show that three CNNs selected in this thesis can effectively recognize the modulation type of the primary signal,and the network of Inception V1 performs slightly better.