Research On Modulation Recognition Technology Based On Deep Learning

Modulation recognition technology is one of the key technologies in communication systems.It affects all aspects of people's life and national development and is applied to many fields such as civil and military.The diversification of communication requirements and the complication of the communication environment prompt the modulation methods to be continuously updated,which undoubtedly brings severe challenges to the modulation recognition technology.Identifying the modulation method of the authorized user's signal is the primary task of cognitive radio technology.With the commercial use of 5th-generation(5G),CR as the key technology of 5G,presents new challenges to modulation recognition technology.In recent years,experts and scholars have been committed to the research of modulation recognition technology and achieved remarkable results.The modulation recognition methods can be divided into two categories: methods based on likelihood decision theory and methods based on feature extraction.The first type of method requires a large number of signal parameters and its calculation process is complicated.Therefore,in practical applications,a method based on feature extraction is often used to recognize the modulation method.Because the recognition effect of traditional feature-based extraction methods mainly depends on the accuracy of feature extraction and the rationality of design classification rules.Considering that deep neural networks have strong self-learning capabilities and can more accurately automatically extract the features of samples,the method of deep learning to recognize modulation methods has received widespread attention.This paper focuses on the modulation recognition technology based on deep learning.Based on the introduction of signal modulation technology and the basic theory of deep learning,we first study modulation recognition methods based on in-phase and quadrature(IQ)features and modulation recognition method based on frequency domain features.We analyse the recognition performance of three different types of deep neural networks,including Convolutional Neural Networks(CNN),Residual Network(Res Net)and Convolutional Long Short-term Deep Neural Network(CLDNN).Then,we analyse the impact of sample length on recognition accuracy.After completing the simulation experiments of the above two methods,we improved the implementation steps when training a deep neural network and proposed a training method based on Signal Noise Ratio(SNR)segmentation.We first divide the large range of the signal-to-noise ratio corresponding to the sample into several adjacent,non-overlapping small cells according to a specific rule,then,extract the samples within the small cell range to obtain several small samples,and finally use several small samples to train the deep neural network to get several trained deep neural networks.In practical applications,these several trained deep neural networks are respectively responsible for recognizing the modulation modes of signals in different SNR intervals.Finally,we study modulation recognition methods based on joint features.Based on the simulation analysis of existing modulation recognition methods based on IQ and statistical features,we propose a modulation recognition method based on amplitude and spectral amplitude features.First,we consider both the time domain and frequency domain characteristics of the signal,and combine the amplitude and spectral amplitude characteristics of the signal to obtain samples.Then,we use the samples to train a “two-in-single-out deep neural network” and a one-dimensional convolutional neural network.Finally,by comparing the recognition performance of the trained “two-in-single-out deep neural network” and the one-dimensional convolutional neural network,we determine that the method has the best recognition performance when using a one-dimensional convolutional neural network.Then,through simulation verification,the recognition accuracy of this method is significantly higher than the modulation recognition method based on frequency domain features studied in Chapter 3.The method further analyses the impact of the number of convolutional layers and the number of convolution kernels on the recognition performance of the one-dimensional convolutional neural network.Simulation results show that excessively increasing the number of convolutional layers or the number of convolution kernels can cause fitting problems to reduce recognition accuracy.