Research On Communication Signal Recognition Based On Enhanced Density Constellations And Deep Learning

With the rapid development of wireless communications and their corresponding services in recent years,the types and bandwidths of wireless communications are growing exponentially.However,the development bottleneck lies in limited spectrum resources.Consequently,it brings challenges to the radio spectrum analysis,which aims to optimize the allocation of spectrum resources and improve the utilization rate of radio spectrum.Moreover,automatic modulation classification(AMC)is the key of radio spectrum analysis.In complex electromagnetic environment,its accuracy and stability are difficult to guarantee.Therefore,this paper focuses on the key theories and technical methods of AMC to meet the needs of radio spectrum analysis under low SNRs,aiming to achieve high precision and robustness target by using deep neural network.The main contents of this paper are as follows.From the point of view of the development of wireless communications and business requirements,this paper firstly explains the significance of the research topic,and then deeps into the principle of wireless signal modulation.Moreover,it introduces the signal representations of orthogonal modulations and the problems existing in the current AMC methods.To solve the problems of signal representation and learning ability in existing methods,a novel neural network-based AMC method named CCNN is proposed.The enhanced density constellation images are proposed to improve the signal representation capability.The compressive convolution neural network is designed using compressive loss function based on the separation and aggregation of feature classes.The simulation platform of different modulations under different SNRs is designed.Simulations show that the proposed method can further enhance the signal representation ability,improve the learning ability of the network model,and obtain about 0.5-2dB performance gain at 95% accuracy compared with the traditional deep learning methods.To solve the problems of the range of recognizable modulations and training complexity,a novel AMC method named HCNN is proposed.Deep convolution neural network is constructed with attention mechanism based dense connection network using multi-features input.The construction of HCNN is more suitable for input features.The simulation platform of different modulations under different SNRs is designed.Simulations show that the proposed method could mine the correlation between different channels of signal features,realize the learning ability of features at the global level,expand the recognizable modulations,and ensure the accuracy of signal recognition under low training complexity.