Researches On Digital Signal Modulation Recognition Technology Based On Deep Learning And Feature Fusion

With the increasing complexity of communication scenarios and the diversification of modulation modes,the traditional modulation recognition methods based on likelihood function ratio and machine learning are difficult to meet the high-performance recognition of modulation signals under the conditions of multiple modulation modes and low signal-to-noise ratio.Thanks to the rapid development of deep learning technology,the method of automatic signal feature extraction combined with deep learning has made important progress in the research field of communication signal modulation recognition.However,how to make full use of the existing signal samples and further explore the characteristic information is still the main work to realize the accurate recognition of modulation signals.Therefore,the overall research of this thesis mainly improves the accuracy of signal modulation recognition from three aspects: mining more significant characteristics of distinction between categories,optimizing data sets and designing neural network models with excellent performance.This thesis mainly applies deep learning to the task of modulation recognition of Multiple Amplitude Shift Keying(MASK),Multiple Phase Shift Keying(MPSK)and Multiple Quadrature Amplitude Modulation(MQAM)signals.The method of feature fusion is used to deeply mine the feature information with more significant difference between categories,and two schemes are proposed.Scheme 1: from the perspective of neural network model,this thesis integrates the multiscale features of shallow layer and deep layer to enhance the expression ability of the model and improve the accuracy of modulation signal recognition.Firstly,the communication signal is mapped into a constellation map,and the feature is enhanced according to the image enhancement technology to build a data set with strong separability of data samples.Secondly,the neural network model takes the classical Visual Geometry Group Network(VGGNet)as the basic architecture,uses the Feature Pyramid Networks(FPN)structure to realize the crosslayer fusion of multi-scale features,constructs the high-level semantic feature maps,and connects Gated Recurrent Unit(GRU)modules in series for sequence modeling.Scheme 2: from the perspective of features,this thesis combines multiple features for feature fusion,and uses the complementarity between features to improve the accuracy of modulation signal recognition.Considering the respective advantages of original In-phase and Quadrature(IQ)signal,high-order cumulant feature,signal-to-noise ratio feature and constellation,feature fusion learning is realized by feature stitching in data set or multi input neural network model.Finally,the experimental results of scheme 1 show that the recognition rate of the enhanced constellation map is 4.36% higher than that of the unenhanced constellation map,and the overall average recognition rate of the designed neural network model is 90.38%.Compared with Alex Net,Goog Le Net,VGGNet and Res Net deep learning models,the recognition rate is significantly improved at low signal-to-noise ratio.The experimental results of scheme 2 show that the recognition method based on feature combination is obviously better than the recognition method based on single feature.The combination learning of features makes the neural network model learn the features with more significant difference between categories.