Research On OFDM Signal Recognition Technology Based On Deep Learning

Automatic modulation classification(AMC)is a key step in signal demodulation,providing a guarantee for the subsequent demodulation work of the receiver.The traditional modulation classification algorithm requires complex pre-processing of the received signal and manual feature mainly based on the statistical characteristics of the signal,with relatively poor robustness;while the modulation classification algorithm using deep learning can reduce the complexity of recognition and achieve good recognition results;among various types of communication signals,the orthogonal frequency division multiplexing(OFDM)signal,as a core of the 4th generation mobile communication technology,is the focus of this thesis,which can be divided into cyclic prefix(CP),cyclic prefix free(CPF)and zero prefix(ZP)according to its guard interval of different filling methods,corresponding to the CP-OFDM signals、CPF-OFDM signals 、 ZP-OFDM signals;This paper proposes an OFDM signals recognition algorithm with a two stage network structure,which can flexibly modify the network structure according to the importance of each level of network recognition task,and solve the problem of easy confusion between OFDM signals and noise under low signal to noise ratio.The details of the research are as follows:First,this paper introduces the basic theory of OFDM signals and compares the differences between traditional multi-carrier modem systems and OFDM modem systems;introduces the foundation of deep learning networks,focusing on the introduction of convolution neural network(CNN)and residual network(ResNet);on this basis,an OFDM signals recognition scheme based on two stage neural network structure is proposed.Secondly,the first stage OFDM inter-class recognition network in this paper is used to identify three types of targets: multi-carrier OFDM signals,single-carrier signals and noise set.The pre-processing methods for non-cooperative signals mainly include coarse estimation of signal bandwidth,variable rate low-pass filtering,time-frequency domain conversion with Welch method and short-time Fourier method,also the performance simulation analysis is carried out for the coarse estimation method of bandwidth,and the estimation results are the precondition for variable rate low-pass filtering.The inter-class recognition performance of different input paths in network is compared,including time sequence channel,frequency sequence channel and time frequency image channel,among which the time frequency image input has the best recognition performance because it includes both time and frequency domain information.The proposed inter-class recognition algorithm is CNN based on single path network,which can filter out the single-carrier signals,and the complexity is significantly lower compared with the ResNet based on single path network.Thirdly,the second stage OFDM intra-class recognition network in this paper mainly distinguishes four types including CP/CPF/ZP-OFDM signals and noise set.The output of the first stage network has some residual noise,which needs to be eliminated in the second stage network.Above all,the ResNet based on dual paths network structure combine the dual input channels of frequency domain sequence and time frequency image,focusing on identifying three types of OFDM signals,and the frequency domain sequence channel can be seen as a supplement to the time frequency image channel information;in order to further improve the network performance,a ResNet network based on multi feature fusion model is proposed,where the artificial features are selected mainly based on the cyclic spectrum and the higher-order accumulation of the signals.The simulation results show that the overall recognition performance is improved compared with the network structure without multi feature fusion,especially at low signal to noise ratios.Fourthly,this thesis proposes an OFDM signal recognition algorithm based on two stage network structure,and conducts a comparison test between the OFDM recognition algorithm with a single stage network structure and the OFDM recognition algorithm with the two stage network structure.The results of the simulation experiments show that the average recognition rate of the cascaded structure network is improved,and the overall recognition rate is better as it improves the confusion caused by noise under low signal-to-noise ratio.