Research On MQAM Modulation Signal Recognition Algorithm Based On Convolutional Neural Network

With the rapid development of society,people have higher and higher requirements on information transmission rate,which makes the limited spectrum resources become more strained.Multi-level quadrature amplitude modulation(MQAM)has high spectrum efficiency and is widely used in the fields of automatic modulation and coding technology,which can greatly alleviate the problem of spectrum resource shortage.In the application of adaptive modulation and coding technology,in order to reduce signaling overhead and improve transmission efficiency,automatic modulation mode identification(AMC)is often carried out at the receiver.Therefore,the MQAM modulation type recognition has very important research significance.In recent years,deep learning tools,such as Convolutional Neural Network(CNN),have been widely used in various fields owing to their excellent deep mining ability.Many researchers also apply them to modulation type recognition and get good recognition results.Therefore,this thesis will study MQAM modulation signal recognition based on CNN.First of all,traditional constellation cannot reflect the density information of overlapping constellation points,and graphic constellation projection(GCP)algorithm cannot fully reflect the position information of constellation points.This paper proposes a gradient color constellation diagram(GCC)algorithm based on constellation density,which converts the density information into color information without destroying the position information of constellation points to realize visualization.Seven kinds of MQAM signals,including 4QAM,8QAM,16 QAM,32QAM,64 QAM,128QAM and256 QAM,are identified and classified by the deep learning ability of CNN.The experimental results show that the recognition rate of this algorithm is 3%-4% higher than that of GCP algorithm at low SNR region.Besides,deep CNN cannot improve the classification performance by deepening the number of network layers,and softmax loss function cannot describe the more discriminative deep features of sample data.This paper proposes a method for classifying seven MQAM signals by combining the residual network with the angular softmax loss function,using a deeper residual network as the main architecture.Angular softmax loss function is used as the output part and the monitoring part of the network architecture to increase the distance between the features of classes and characterize more discriminative features.Finally,a series of comparative experiments are done to prove the effectiveness and superiority of the proposed algorithm.