Deep Learning Based M-QAM Modulation Signal Recognition

With the rapid development of communication technologies,the modulation methods of signals have become diversified and complicated.In most cases,the receiver does not know the modulation information of the signals.Therefore,how to automatically recognize the modulation modes of the received signals becomes particularly important.MQAM signal has high spectrum resource utilization,so it has been widely used in wireless broadcasting and other fields.Because the shallow learning method can't extract the deep features of the signal and can't solve the complex problem effectively,moreover,constellation features can't effectively express the features such as overlapping signal points and denseness.This paper uses GCP algorithm to extract the image features of the signal which is used as the input to the CNN,then the paper mainly studies the MQAM signal in the gaussian noise channel,the main research contents are as follows:1.The softmax loss in deep learning can only make the deep features separable but doesn't take into account the distribution of them.To address this issue,this paper combines the center loss metric learning and softmax loss,and proposes a new method called MQAM recognition method based on deep metric learning.The center loss metric learning can de-scribe the intra-class dispersion of the deep features.With the joint supervision of center loss and sofftmax loss,not only the inter-class features differences can be enlarged,but also the intra-class features variations can be reduced.This paper has carried out sufficient ex-periments on the seven modulation signals of 4QAM,8QAM,16QAM,32QAM,64QAM,128QAM and 256QAM.The results show that the proposed method can improve the classi-fication performance of MQAM signals.2.Convolutional neural network based on deep metric learning can't effectively select fea-tures.To address this shortcoming,a MQAM signal recognition method based on deep metric learning and feature selection has been proposed.It models linear and nonlinear global descriptors of feature channels and uses the modeled results as weights to measure the dependencies and importance of feature channels.It can selectively emphasize useful features and suppress less useful features finally.This paper has carried out sufficient exper-iments on seven kinds of QAM signals.The experimental results prove the effectiveness of the proposed algorithm.