| The digital modulation recognition is located at the receiving end of the digital communication system,and the modulation type identification of the signal is required before the signal is demodulated.The identification of the digital modulated signal is to extract some characteristics of the signal at the receiving end,and select a suitable classifier to classify and identify the signal.According to the identified modulation mode,the corresponding demodulator is selected for demodulation,thereby recovering the original signal.The purpose of this paper is to use the instantaneous BP signal,the high-order cumulant,and the constellation feature as the modulation signal characteristics for four common and ten common digital modulated signals under different SNR conditions.The road and convolutional neural networks act as classifiers to identify the set of identified signals.The main research method in this paper is to select MASK,MFSK,MPSK and MQAM signals in the selection of the signals to be identified.In the feature quantity selection,the feature quantities used are instantaneous feature quantities,high-order cumulants and constellation features.The classifier is selected using a BP neural network classifier and a convolutional neural network classifier.The simulation platforms used in the experiments were MatLab2014 a,TensorFlow machine learning framework,and python-related libraries.There are a total of five points in the work and innovations included in the paper.Firstly,a BP neural network algorithm based on improved particle swarm optimization(PSO)is proposed to solve the problem of falling into local minima during network training.Secondly,a convolutional neural network algorithm based on the improved LetNet-5 model is proposed for the recognition of the set {4QAM 16 QAM 64QAM},which reduces the network training time.Third,the recognition signal set is subjected to recognition simulation using five common transient signal characteristics,and the verification of the transient signal feature is capable of identifying the set of signals to be identified.Fourthly,based on the fourth-order and sixth-order cumulants,the high-order cumulant feature is constructed and the identification signal set is tested and simulated.The extracted high-order cumulant is able to identify the signal set to be identified.From the final experimental simulation results,first,compared with the traditional algorithm,the BP neural network algorithm based on improved particle swarm optimization improves the recognition rate of each signal.When the signal-to-noise ratio is equal to 5dB,each signal The recognition rate is close to 85%.Second,the convolutional neural network based on the improved LetNet-5 model can better distinguish different QAM signals,and the recognition rate is close to90% when the signal-to-noise ratio is 5dB. |
