Research On Modulation Recognition Algorithm Based On Deep Learning

Automatic modulation recognition is mainly used in non-cooperative communication scenarios such as spectrum monitoring,electronic countermeasures,cognitive radio and adaptive communication.With the rapid development of communication technology,the electromagnetic environment is becoming increasingly complex,which increases the difficulty of signal modulation recognition.The traditional modulation recognition method is difficult to meet the requirements of modern communication.In this thesis,the deep learning technology is applied to the research of modulation recognition algorithm to improve the performance of modulation recognition algorithm in complex electromagnetic environment.Firstly,this thesis studies the performance of three classical neural network models in deep learning(CNN,Res Net,Dense Net)in modulation recognition,and analyzes the influence of network hyperparameters on the accuracy of modulation recognition.The experimental data show that if the above three neural networks are used for modulation recognition alone,the recognition accuracy is not ideal at low SNR.In order to further improve the accuracy of modulation recognition,and considering that one-dimensional convolution neural network is more suitable for processing the signal data of sequence type,this thesis designs a modulation recognition network model DRN based on onedimensional convolution in SISO system by combining Res Net and Dense Net.The model fully integrates the idea of identity mapping of Res Net and the dense connection structure of Dense Net,that is,six residual blocks based on one-dimensional convolution are connected by the dense connection mode of Dense Net to form a new network model.At the same time,neural network optimization methods such as batch normalization,adding Dropout layer and Adam optimizer are adopted to further improve the performance of the model.The experimental results show that the proposed modulation recognition model has good robustness and the recognition accuracy of 93.7%is achieved with low computational complexity.In the non-cooperative MIMO system,because the number of transmitting antennas is unknown and the mixed signals of multiple transmitting antennas are received,the modulation mode of the signal cannot be identified directly by the algorithm in SISO system.In order to solve the above problems,this thesis first estimates the number of transmitting antennas through the Gerschgorin Disks Estimation algorithm,and then uses the fixed-point algorithm(Fast ICA)in the blind source separation algorithm to separate the mixed MIMO signals.The signal separated by Fast ICA algorithm has amplitude error and phase rotation,while the effect of CNN will become worse when the data is rotated or tilted.The Caps Net overcomes the deficiency of CNN to some extent,so this thesis proposes a network model which combines DRN and Caps Net to identify the modulation mode of the separated signal.This model mainly uses the encoder of Caps Net,specifically,the first convolution layer of Caps Net encoder is replaced with the convolution part of DRN,and the whole network is based on one-dimensional convolution.The experimental results show that the recognition accuracy of the algorithm can reach 99.51% under high signal-to-noise ratio,and has good generalization ability.