Research On Automatic Modulation Classification Of Digital Signals Based On Deep Learning

With increasing diversity of the applied digital modulation schemes,AMC(Automatic Modulation Classification)becomes a significant research hotspot.It is urgent to propose a more simplified and more accurate modulation recognition strategy based on recently developed AMC technologies.Standing on present AMC studies,this thesis attempts to lower the burden in signal pre-processing step and investigates the threads for applying deep learning algorithms into this field.For achieving the above-mentioned goals,this thesis carried a series of researches and studies on DL(Deep Learning)algorithms,specifically on DBN(Deep Belief Network)and CNN(Convolutional Neural Network).The main contributions of this thesis are listed as follows:1.A GCP(Graphic Constellation Projection)algorithm is proposed to project signal's constellations into artificially structured images,which are subsequently sent to DBN-based classifier for feature learning and modulation classification.This scheme is inspired by the powerful image recognition ability of DBN and its performance superiority to other schemes is experimentally demonstrated by simulation results.2.An ASFPF(Amplitude Spectra Feature and Phase Feature)and 1D(one-dimensional)CNN based AMC algorithm is proposed.This algorithm involves a two-cascaded CNN classifier,which can dig out the modulation scheme information hidden in simple inputs.In this whole classifier,the first CNN aims to classify signals with different amplitude spectra features,while the following cascaded one can distinguish signals with different phase features.3.A modulation classification method that can classify various modulation patterns is proposed.This method carries out Hilbert orthogonal transformation on the received signal at first and then sends the in-phase and quadrature components of the transformed signal to CNN classifier for modulation classification.CNN extracts the characteristics of the signal in each layer and get the modulation pattern according to the features of the known signals.Compared with the conventional AMC methods,which requires a large number of calculations to obtain the signal feature parameters,the modulation classification of digital signals based on DL is simple and the simulation results show that it has high classification accuracy at low signal-to-noise ratio.