Modulation Recognition Algorithm Based On Spectral Feature Extraction

Modulation classification of communication signals is a key part of non-cooperative signal processing.Especially,the method based on statistical pattern recognition is the current research hotspot.Due to the lack of prior information such as signal frequency and phase,as well as the influence of external environment noise on the stability of signal features,the modulation recognition of spatial signals has always been a difficult problem.This article introduces the method of spectral feature extraction with strong anti-noise capability and less prior information.Then we discuss the method of classification.This paper analyzes various types of spectrum of digitally modulated and analog modulated signals,including the power spectrum,high-order spectrum,envelope spectrum and fractional low-order cyclic autocorrelation spectrum.The algorithm of spectral line detection based on spectrum refinement is proposed.The influence of "fence effect" of discrete spectrum on the detection value is eliminated by using the algorithm.Features are extracted by detecting the spectral line.The extraction process uses the method of spectral region division and requires less prior information.Modulation recognition processes based on decision tree and vector matching are designed respectively.This paper discusses the problem of setting the optimal threshold for the known and unknown SNR parameters: Under known parameters,we use Bayes method to determine the optimal threshold.Under unknown parameters,the application scenario needs to be selected first,then the feature probability density is estimated from the samples,and the optimal threshold is determined using the minimum risk method.Finally,it realizes the scheme of modulation recognition with high recognition performance and strong antinoise ability.In the end,the modulation recognition method based on supervised learning is studied.The support vector machine and error correction output coding method are applied to modulation recognition.Spectral features are used as input data.We discussed the classification performance under four types of coding matrices of “one-versus-all”,“oneversus-one”,“dense random” and “sparse random”.We compare the performance of three schemes of decision tree,vector matching and multi-class support vector machine from two aspects of simulation experiments and measured satellite signal.The results show that the multi-class SVM classification method is generally better than the decision tree and vector matching methods.Under the condition of data in this paper,the recognition probability approaches to 100%.