Research On Digital Modulation Recognition Based On Neural Network

With the increasing diversification of communication signal modulation methods,communication signal modulation pattern recognition plays an important role in the field of signal analysis.Due to the influence of channel environment and modulation order,the traditional modulation recognition method based on likelihood detection and feature extraction is less accurate.The neural network has achieved excellent performance in speech recognition and image recognition,and has strong generalization ability and robustness.Therefore,it is feasible to apply neural network to modulation recognition.Traditional neural network identification methods usually take signal characteristics as input.This method is greatly affected by the quality of the input features and cannot be extended to other modulated signals.Recently,some researchers have used deep neural networks to directly identify modulated signals and achieved good recognition results.Therefore,this paper deeply studies the identification method of digital modulation based on neural network.The main work is summarized as follows.This paper proposes a digital modulation recognition method based on One-Dimensional Convolutional Neural Network(1-D CNN).This method designs a 1-D CNN classification model for the one-dimensional sampling sequence of the IF modulated signal,and uses the known modulation sampling sequence and modulation method to supervise the network,and then uses the trained network to identify the modulated signal.In order to obtain the best 1-D CNN recognition performance,a set of control experiments was designed to determine the optimal parameters of the network,including sample length,training set and other factors.The experimental results show the 1-D CNN modulation identification method proposed in this paper achieves good recognition performance for 2ASK,2FSK,4FSK,8FSK,2PSK,4PSK and 8PSK signals.When the signal-to-noise ratio is greater than 5d B,the recognition rate of this method exceeds 99%..And the average recognition rate of the method exceeds 58% when the signal to noise ratio ranges from-20 d B to 0 d B.Aiming at the problem that 1-D CNN has low recognition rate for some ASK signals and QAM signals,this paper proposes a digital modulation recognition method based on Deep Residual Network(DRN)and Curriculum Learning(CL).This method uses a deeper depth residual network to achieve better network recognition performance,and combines course learning to reduce the impact of signal noise on network recognition performance.Similar to 1-D CNN,this method also does not require traditional pre-processing steps to directly identify the IF-modulated signal.The difference is that the method converts the onedimensional sample sequence into a gray-scale image by simple data conversion,and then import grayscale images into the classification network.In the training phase,this paper uses a neural network called the Mentor Net to provide a curriculum to supervise the training of the DRN.In order to help the DRN learn a more robust network model.During the testing phase,the DRN works independently to identify the input modulation signal.In this paper,the performance of the identification method is verified in an environment containing additive white Gaussian noise,multipath fading,carrier frequency offset or phase offset.The experimental results show that the proposed method can achieve a recognition rate of 99.3% for 11 commonly used modulation types when the signal-to-noise ratio is high,and the interclass precision can reach 100%,which is much higher than other identification methods.In addition,in the presence of interference,the overall accuracy of the identification method can reach 90% under the signal-to-noise ratio of 10 d B,which has good robustness and is suitable for applications such as military electronic warfare.