Research On Modulation Recognition Of Digital Signal Based On Multi-feature Extraction

Modulation classification that is the first step in obtaining signal information plays a key role in various military and civilian applications,such as electronic warfare and interference cancellation.For many years,with the diversification of communication signals and the complexity of the channel environment,higher requirements have been placed on modulation recognition technology.This paper will study the feature-based modulation methods with machine learning.In this paper,twelve possible modulation categories are considered,including{2ASK,4ASK,8ASK,2PSK,4PSK,8PSK,2FSK,4FSK,8FSK,16QAM,32QAM,64QAM}.Based on the instantaneous information,cyclostationarity and high-order cumulants,we propose that we achieve inter-class identification firstly,then further intra-class identification,and verify the resolution of each feature parameter to the signal under different signal-to-noise ratios with 8 featurers according to the difference of different dimensions of the signal And we proposes an effective feature parameter—the number of first-order cyclic moments of the signal to accomplish FSK classification.In order to solve the problem of spectral peak submergence during feature extraction,the segmented autocorrelation accumulation method is used.Improvements have been made to significantly enhance the spectral peaks.Experimental results show that FSK's intra-class recognition can achieve 99%accuracy when the signal-to-noise ratio is higher than-5dB.We use decision tree to accomplish modulation classification based 8 extracted multi-dimensional feature parameters.And we explore the impact of depth on accuracy and stability,and verify the perfect performance of the CART algorithm when SNR is higher.Then Adaboost integrated algorithm based on decision tree is used to make up the decision tree model,improving the stability and generalization ability.Lastly,through a comparative analysis of weak classifiers,a support vector machine(SVM)with stronger representation ability was selected as the base classifier of Adaboost,improving recognition accuracy under low signal-to-noise ratio.Particle swarm algorithm was used to optimize the hyperparameters of SVM.The results prove that Adaboost-SVM has excellent recognition performance under low signal-to-noise ratio.When the signal-to-noise ratio is higher than-3dB,the signal recognition accuracy rate can reach more than 95%.