Research On Automatic Modulation Recognition Technology Based On Deep Learning

Most traditional methods for automatic modulation recognition first extract features from modulation signals,and then classify different modulation types by these features with machine learning algorithms.The performance is limited by how to extract and select effective features.These methods are hard to adapt new unknown modulation signals.The time domain samplings nearly contain all information of modulation signals which are hard to be directly utilized by traditional methods.Deep learning algorithms have dominated the area of image recognition and speech signal processing in recent years,especially these algorithms process original image data or speech data directly.Therefore this thesis will discuss how to apply deep learning in automatic modulation recognition,the main studies are as follows:Firstly,this thesis proposed one time-domain modulation recognition algorithms SCC-CNN(based on CNN)for general communication modulation signals.The performance of deep learning based methods improved about 3～5dB than traditional methods,no expert feature engineering is needed at the same time.Secondly,this thesis proposed another time-domain modulation recognition algorithms DC-LSTM(based on LSTM)for general communication modulation signals.The performance of this method improved about 0.5～1dB than former methods,this algorithm has comparable performance compared with SCC-CNN.Thirdly,it proposed a deep learning features represented algorithm for open set modulation recognition which was the first effort in this area.This algorithm learned discriminative features by deep learning,then the distribution of features ware fitted and used to achieve open set modulation recognition.The open set modulation recognition performance of two different data set improved 14.2%and 24.4%respectively.Finally,a simple algorithm-CB-CNN was run on an FPGA platform after fixed-point compression.