Research On Signal Recognition Technology Of Non-cooperative Communication

In recent years,with the rapid development of the communication industry,the modulation recognition technology in non-cooperative communication systems has been paid more and more attention,which has become a research hotspot of many scholars.Therefore,this paper conducts in-depth research on signal modulation recognition technology in non-cooperative communication.The specific research work and results are as follows:1.A modulation identification method based on higher-order cumulant and fractionalorder wavelet transform is proposed.In this method,the signal analysis tool fractional wavelet transform is applied to modulation recognition.Firstly,the received signal is decomposed by seven layers of fractional wavelet,and then the detail components of each layer of wavelet coefficients are added to construct the feature parameters,which are combined with the fourth-order and the sixth-order cumulant features,and the decision tree classifier is constructed to identify the signal by setting the decision threshold.According to the simulation analysis,when the SNR is 4d B,the recognition rate reaches more than 90%,and when the SNR is greater than 8d B,it is close to 100%.In the range of SNR between-10 d B and 0d B,compared with the other three recognition methods that combine features such as higher-order cumulant and wavelet transform,the method in this paper has the highest recognition rate,which proves that the features based on fractional wavelet transform have good noise resistance.This method completes the classification and recognition of ten kinds of digital modulation signals with only three characteristic parameters.2.A modulation recognition method based on SSA-ELM(Sparrow Search AlgorithmExtreme Learning Machine)is proposed.This method is an improvement of the above decision tree recognition method.Considering that the decision tree classifier needs to set the decision threshold manually,the extreme learning machine is selected as the classifier.At the same time,in order to further improve the signal recognition rate under low signal-tonoise ratio,the SSA algorithm is used to optimize the parameters of the ELM.The simulation results show that compared with the unoptimized classifier ELM,the recognition performance of SSA-ELM is significantly improved.Under the low signal-to-noise ratio of-10 d B ～ 0d B,the average recognition rate of SSA-ELM is 30.4% higher than that of decision tree.Compared with the traditional genetic algorithm,it is also proved that SSA has better optimization performance.In addition,the three recognition methods based on SSAELM,ELM and decision tree are respectively verified by the measured signal data collected by USRP.3.An end-to-end recognition network based on CBDDNN(Convolutional,Bidirectional Long Short-Term Memory,Deep Residual Shrinkage Module,Deep Neural Networks)is proposed.Firstly,the one-dimensional convolution is combined with the LSTM(Long Short Term Memory)network and its two variants: GRU(Gated Cycle Unit)and BiLSTM(Bi-directional Long Short-Term Memory)to construct three networks of CNNLSTM,CNN-GRU and CNN-Bi LSTM,using simulation data set and public data set RML2016.10 a verify the recognition performance of the three networks respectively,the highest recognition rate of CNN-Bi LSTM is more than 91%.Then,in order to further improve the recognition rate of low SNR,the DRSM(Deep Residual Shrinkage Module)with soft threshold noise reduction function is added to CNN-Bi LSTM to obtain the improved CBDDNN.In the SNR ranges from-10 d B to 0d B,the recognition rate of CBDDNN in the simulation dataset and RML2016.10 a dataset is 0.29% and 2.63% higher than that of CNN-Bi LSTM respectively.When CBDDNN is used to classify USRP measured signals,the highest recognition rate reaches 96.5%.