Research On Modulation Pattern Recognition Under Low SNR

Digital signal has a high utilization rate of spectrum,can better resist channel loss,and has excellent anti-noise performance and good confidentiality.It can reuse various forms of information such as voice data and video images.Digital modulation is increasingly widely used in modern communication.Compared with M-ASK and M-FSK,M-PSK has the advantages of high transmission efficiency,strong anti-noise performance,and less affected by the change of channel characteristics.It is widely used in high-speed data transmission.Multi-ary quadrature amplitude modulation(M-QAM)is an important new digital band-pass modulation technology,which can guarantee the bandwidth and power of multi-ary phase shift keying(M-PSK),and improve the disadvantage of low noise tolerance of PSK when the number of bits increases.Therefore,this dissertation mainly focuses on two excellent and widely used phase shift keying(M-PSK)modulation methods and orthogonal amplitude modulation(M-QAM)modulation to carry out modulation recognition under low SNR.In this dissertation,the experimental scheme of pattern recognition with low signal-tonoise ratio is optimized through the experimental design.Through the noise reduction preprocessing of the received noisy signal,the signal characteristics with strong representation ability are determined by time-frequency analysis according to its modulation principle.Based on the improved Res Net50 modulation recognition model,modulation recognition research is carried out on eight modulation signals,including BPSK,QPSK,8PSK,16 PSK,4QAM,16 QAM,32QAM and 64 QAM,at low SNR.The main research contents and innovations of this dissertation include the following three aspects.(1)Digital signal denoising based on Denoising Auto-encoder(DAE).Aiming at the problem that the recognition accuracy of low signal-to-noise ratio signal modulation is low and the effect of traditional denoising algorithm for noisy signal needs to be improved,this dissertation makes the following improvements: The denoising autoencoder algorithm based on convolutional neural network is used to denoise the noisy signal.The denoising effect of BPSK,QPSK,8PSK,16 PSK,4QAM,16 QAM,32QAM and 64 QAM signals is compared with the traditional wavelet threshold denoising algorithm under the condition of-5 d B to 15 d B interval of 2.5 d B SNR.Taking the signal error rate parameter as the evaluation index,th e simulation experiment verifies that the noise reduction effect of the noise reduction selfencoder algorithm introduced in this dissertation is better than that of the traditional wavelet threshold denoising algorithm,which improves the accuracy of modulation pattern recognition under low SNR conditions.(2)Digital signal feature extraction based on time-frequency analysis.Aiming at the problems of insufficient time-frequency analysis and single feature extraction of digital signals,this dissertation makes the following improvements: Based on the modulation principle,the time-frequency analysis selects the signal spectrum with strong representation ability of constellation and cyclic spectrum as the data characteristics of M-PSK and M-QAM modulation signals.After Z-axis normalization of the cyclic spectrum is converted to RGB cyclic spectrum,the constellation diagram and RGB cyclic spectrum are further converted to grayscale map for data superposition as the input data of the network.In the aspect of data preprocessing,the signal features extracted after video analysis are verified by comparative experiments as the input data of the network model after data superposition,which effectively improves the accuracy of modulation recognition.(3)Digital modulation pattern recognition algorithm based on improved Res Net50 feature splicing.Residual neural network performs well in classification task.In this dissertation,Res Net50 network model is selected for classification based on deep learning modulation pattern recognition research scheme.In order to solve the problem that the single feature of input data leads to the low accuracy of signal classification,the signal constellation gray scale map and the signal RGB cyclic spectrum gray scale map are superimposed as the input of the network in this dissertation.The statistical information between each channel of the feature map is extracted by embedding the channel attention module of Res Net50 to complete the recalibration of the data characteristics of each channel of the feature map and realize the optimization of the network model.In this dissertation,the simulation of M-PSK,M-QAM digital signals by adding Gaussian white noise to simulate the actual communication of-5d B to 5d B(interval 2.5d B)noise signal,after denoising autoencoder noise reduction processing converted to signal constellation gray scale and RGB cyclic spectrum gray scale,through feature superposition as the input data of the improved Res Net50 model,modulation pattern recognition.Through the comparative experiments of three simulation data sets based on the improved Res Net50 network model,the results show that the recognition accuracy of the simulation data set after the superposition of the two data is higher.By comparing the accuracy of different convolution network modulation pattern recognition based on the same data set,the results show that the improved Res Net50 modulation pattern recognition is better.The experimental results show that under the condition of low signal-to-noise ratio,the research scheme designed in this dissertation has better accurate recognition effect for eight kinds of M-PSK and M-QAM digital signal modulation modes.