Time-frequency Analysis And Application In Modulation Recognition Of Low Snr Signal

In a complex electromagnetic environment,all kinds of modern electromagnetic equipment radiate a lot of electromagnetic signals.There are many modulation methods of these electromagnetic signals.Accurately identifying the modulation method of the signal can provide an effective basis for the next analysis and operation.At present,the method of according to the characteristics of the signal to identify is commonly used.However,the interference of noise has a great impact on the time domain or frequency domain characteristics of the signal.The time-frequency characteristics of signal are less disturbed by noise,and the features are obvious and stable.Therefore,in the case of low SNR,the method of using signal time-frequency characteristics for modulation recognition has recognition advantages.Based on the time-frequency analysis to obtain the time-frequency diagram of the signal,this paper conducts research from two aspects:feature extraction methods are designed by artificial and using deep learning to extract features automatically.The main work is as follows:(1)Some time-frequency analysis methods have been studied.Through performance analysis,Choi-Williams distribution with better performance in terms of time-frequency resolution and cross-term suppression is selected to obtain the time-frequency diagram of the signal.Through the Choi-Williams distribution,a time-frequency map of the signal with clear outline and obvious characteristics can be obtained.(2)The denoising methods of the time-frequency diagram have been studied.Through the analysis of denoising effect,the denoising method by total variation is selected to process the time-frequency graph.After denoising,the signal-to-noise ratio is improved by 2.24 d B.(3)A recognition method based on the corner feature and straight line feature of the signal time-frequency graph is proposed.Harris corner detection algorithm is used to extract corner features of time-frequency graph,and the false corner points will be eliminated.Hough transform is used to extract straight line features of time-frequency images.Combined with corner points and straight line features,the six-dimensional feature vector is obtained.Then input the feature vector into the support vector machine for recognition and classification.The average recognition rate of these six modulation signals for 2ASK?2FSK?4FSK?MSK?BPSK?QPSK is 93.3% under the condition of 0d B SNR.Compared with the recognition method based on other features,the method has better recognition effect under the condition of lower SNR.(4)A method to recognize the time-frequency diagram of different modulation signals by using Deep Belief Network and Convolutional Deep Belief Network which are improved The gray-scale time-frequency diagrams and the binary time-frequency diagrams of six kinds of modulation signals are input into these two deep learning models for extracting features automatically.Then the deep learning model can training and recognizing these diagrams.Under the condition of 0d B SNR,the Deep Belief Network has a good recognition effect on the six kinds of binary time-frequency graphs,with a recognition rate of 99.4%.The Convolutional Deep Belief Network can effectively identify both gray-scale time-frequency images and binary time-frequency images,and the recognition rates are 94.7% and 95.9%,respectively.Under the condition of-4d B SNR,for the recognition of binary time-frequency images,the recognition accuracy of the Deep Belief Network can still reach 96.7% and that of Convolutional Deep Belief Network is 90.2%.Therefore,this method can achieve good recognition results under low SNR.