Study On Application Of Deep Learning Method In Communication Modulation Recognition

The formal commercialization of 5G has promoted the further development of the communications industry.At the same time,the communication environment has become more complex and signal modulation systems have become more diverse.Terminals of communication systems receive signals and demodulate signals that carry information.In this process,the identification of modulated signal systems is important Research direction.At present,there are two main types of methods for modulation signal recognition,recognition methods based on decision theory,and machine learning classification methods based on feature engineering.These two types of methods have perfect theoretical foundations but require sufficient prior knowledge and complicated theoretical derivation.They are greatly affected by the channel environment and the recognition rate is not high under low signal-to-noise ratio.Therefore,this paper studies the identification of modulation signal systems based on deep learning.This method does not require human participation or data statistics.It can automatically extract features from the original data and identify multiple types of modulation signal systems.The main content of this paper is as follows.First,this article summarizes the theoretical basis of deep learning,including the basic neuron structure and its corresponding forward and backward propagation algorithms.In subsequent articles,convolutional neural networks,recurrent neural networks,and fully connected neural networks are mainly used.Therefore,these three types of networks are introduced mathematically.Secondly,this paper studies the popular deep learning models in other fields,implements some models and successfully applies them to the modulation system identification field,and proposes a CLDNN end-to-end network model suitable for modulation signal identification.This network model Compared with other deep learning models,it has higher recognition efficiency and recognition accuracy.The recognition method is subsequently compared with the traditional machine learning-based recognition method to verify the efficiency of the method.The experimental results show that the proposed network model can identify 11 types of digital and analog modulation methods at the same time,and the recognition rate is improved compared with the existing methods under low signal-to-noise ratio.When the signal-to-noise ratio is above-4d B,the average recognition rate for multiple types of modulated signals is about 95%.Then,the influence of network hyperparameters on the recognition performance of CLDNN is studied,including the number of different convolutional layers,the size and number of convolution kernels,and the impact of long-term and short-term memory layer networks.Finally,a series of relatively ideal network hyperparameters were determined.Finally,a modulation identification method for communication signals is designed.Deep features extracted by CLDNN are combined with traditional artificially extracted features such as spectral features and high-order cumulant features to form feature vectors,which are then normalized for classification using an integrated learning classifier.The experimental results show that the modulation recognition method based on feature fusion further improves the recognition accuracy and is superior to the previous method.