Research On Modulation Type Recognition Algorithm Based On Deep Learning

Modulation type recognition is an important link between signal detection and signal demodulation,and it is also one of the key technologies of software radio and non-cooperative communication.There have been a lot of research results on modulation type recognition using maximum likelihood test theory and type recognition theory.In recent years,deep learning has been applied to various fields due to its excellent feature extraction capabilities.Some researchers have applied deep learning methods to modulation type recognition,and the feature parameters are usually used as input data.However,these methods are affected by the feature parameters and have poor expandability.Recently,foreign researchers have proposed a modulation type recognition method that directly uses signal data and residual neural network to obtain better recognition performance.Therefore,this paper deeply studies the problem of modulation type recognition based on deep learning.The main work is summarized as follows.Firstly,modulation type recognition algorithm which use Hybrid-Restricted Boltzmann Machine neural network is proposed.For recognition of commonly used modulation types such as MPSK,MQAM,and MAPSK,high-order moment formula for MAPSK signals are derived in detail based on the signal model,and the theoretical values of each order cumulant of 16 APSK,32APSK,MQAM,and MPSK signals are calculated.Two high-order cumulant feature parameters of BPSK,QPSK,8PSK,16 QAM,32QAM,16 APSK,and 32 APSK signals are obtained according to the eighth,sixth,fourth,and second order cumulants.And then the Hybrid-Restricted Boltzmann Machine neural network is used as a classifier for identification(named FHRBM algorithm).The simulation results show that the correct recognition rate of HRBM algorithm is high.Secondly,a modulation type recognition algorithm based on sparse auto-encoder and feature parameters(named FSAE algorithm)is proposed.The above two high-order cumulant feature parameters are coded as sparse auto-encoder's input samples.The depth features of the coded feature parameters are extracted by the auto-encoder,and then the signal classification is completed by the Softmax classifier.The simulation results show that when the signal to noise ratio(SNR)is greater than 0d B,the average correct recognition rate of SAE algorithm is higher than 90%,and the recognition performance is good.Thirdly,because the FSAE algorithm needs artificially pre-designed feature parameters,when modulation types to identify are different,the feature parameters need to be redesigned and the expandability of algorithm is poor.In order to solve the problem,a new modulation type recognition algorithm based on signal data and sparse stack auto-encoder(named SSSAE algorithm)network is stadied then.The in-phase component I(t)and quadrature component Q(t)of the received signal are used as the original sample data.After vectorization and normalization,the data is used as the input vector of the SSAE.The network training adopts the method of layer-by-layer pre-training and overall fine-tuning.The simulation results show that the SSSAE algorithm can extract more robust features,and the recognition performance for complex MAPSK signal model is good.The disadvantage is that the algorithm has poor recognition performance at low SNR.Finally,in order to further improve the recognition performance based on signal data and deep learning algorithm at low SNR,a convolutional neural network(CNN)with strong feature extraction capability is used to modulation type recognition.The I(t)and Q(t)components are formed into an input matrix with a dimension of 2?128,and a CNN network model and its corresponding convolution kernel are designed.Then,feature extraction and modulation type recognition automatically are performed by the CNN network.The simulation results show that the convolutional neural network has strong feature extraction ability and good recognition performance at low SNR.At the same time,it does not need to extract features manually,and the algorithm has good extendibility.