Research On Modulation Recognition Technology Based On Multi-classification Combination Strategy Optimization

As a key technology in non-cooperative communication system,the main research content of modulation recognition is to determine the modulation mode of the detected radio signals.It has been widely used in the electromagnetic supervision of civil spectrum by government and the electronic warfare in the military field.With the development of communication technology,the modulation modes have gradually become diversified,which puts forward higher requirements for the research of modulation recognition.The main research method of this paper is to use the statistical pattern recognition method based on feature extraction to complete the classification of ten kinds of digital signals,including 2FSK,MSK,2ASK,4ASK,2PSK,4PSK,8PSK,16QAM,32QAM and 64QAM.according to the modulation principle,cyclic spectrum and high-order cumulant theory of the above ten signals and two core technologies of the method,this paper studies the feature extraction and multi-classification system based on combination strategy optimization.First of all,in the aspect of feature extraction,this paper analyzes the differences among12 features such as cyclic spectrum,high-order cumulant and instantaneous information statistics of signals for different modulation modes,and establishes a feature space containing10 kinds of signal samples to be identified from the perspective of multi-dimensional feature extraction.Among them,in order to recognize MSK signal and classify 16QAM and 64QAM signals,this paper constructs two kinds of featuresY_{1 3} andY₁₂ which reflect the statistical characteristics of the envelope amplitude in zero-frequency cross section of the cyclic spectral density function,and two kinds of combined featuresF_{Q 1} andF_Q2 which are based on the principle of higher-order cumulants.And,the influence of several common factors on it is analyzed.Then,in the research of tree multi-classification system design,this paper uses 12 kinds of feature parameters to establish the traditional decision tree classifier model,and completes the initial classification task of ten kinds of signals.And,in order to improve the classification performance of decision tree classifier,support vector machine and feature selection(FS)idea are introduced to design the multi-classification system based on tree shaped smooth support vector machine combined with feature selection algorithm(FS?DT-SSVM).At the same time,the experimental results of the classification system using three different node feature subset selection methods are compared and analyzed.Finally,in the design of parallel multi-classification system,multi-classification system based on confidence weight matrix(CWM)is constructed according to the basic principle of Bayesian formula,which can further enhance the conversion rate of classification performance from basic binary classifier to multi classification system.Moreover,the support vector machine multi-classification system model combined with CWM(CWM?SVM)is established.And then,a comparative experiment is conducted between CWM?SVM and OVO?SVM multi-classification system to verify the classification performance of the proposed algorithm.After that,through the overall analysis of the multi-classification accuracy of ten kinds of digital signals compared with the two classical ensemble learning,the classification performance of CWM?SVM is better than the other three multi classification systems.