Technology Of Digital Modulation Signal Recognition Under Large Dynamic SNR

At present,there are some unresolved problems in digital modulation recognition technology,such as the limited recognition type,the low recognition rate at low signal-to-noise ratio(SNR)or when SNR varies widely.Therefore,it is still a challenging research topic to increase the recognition types and improve the recognition rate based on the original recognition technology.In this paper,the key technologies and theories of digital modulation recognition methods and algorithms are studied,and the main research results are as follows:1.The fractal feature of digital modulation signal is studied deeply,and the one-dimension fractal feature set is constructed to increase the distinguishability of features.Aiming at the limitation that the traditional grey correlation classifier can only handle single values and weights set artificially cannot be changed,the feature is expanded from single value to interval.By updating the weights in real time through the feature data,an adaptive weight interval grey correlation classifier is proposed,which effectively improves the accuracy of digital modulation recognition.2.The entropy features of digital communication signals are studied deeply,and the extraction method of time frequency Rényi entropy and energy entropy is proposed.Sixteen kinds of entropy feature of digital communication signals are summed up to construct the original entropy feature set.In order to solve the problem that there is redundancy in entropy features,the feature selection techniques such as Sequential Forward Selection(SFS)algorithm,Sequential Floating Forward Selection(SFFS)algorithm and ReliefF algorithm are proposed to select the most distinguishable feature subset from the original feature set,which reduces the redundancy of feature subsets.On this basis,Gradient Boosting Decision Tree(GBDT)classifier and Xgboost classifier are proposed to effectively improve the accuracy of digital modulation recognition.3.The classifier based on evidence theory is studied deeply.Aiming at the conflict evidence,a new reliability weighted fusion model is proposed,which effectively improves the fusion effect of the conflict evidence.According to the characteristics of the entropy of the digital modulation signal,the method of the basic probability assignment(BPA)functionacquisition based on Bell-Shaped Function is proposed,which effectively improves the effect of BPA acquisition.In view of the problem of large variation of entropy features when the SNR is low and the SNR varies in a large range,a multi-feature reliability weighted fusion model based on evidence fusion model is proposed,and a classifier reliability weighted model based on SNR estimation is established,which greatly improves the accuracy of digital modulation recognition.4.The deep learning theory,architecture,and commonly used models are studied deeply.A novel concept of data set called contour stellar image is proposed,which increases the separability of the data sets and greatly improves the accuracy of digital modulation recognition.The application effects of deep learning models such as AlexNet,GoogleNet,VGG and ResNet on digital modulation recognition are analyzed,and a data enhancement method based on GAN model is proposed,which improves the accuracy of digital modulation recognition.