Study On Modulation Recognition Based On Constellation Clustering And Artificial Neural Network

The recognition of modulation types of communication signals is an important and fundamental research subject in modern electronic warfare. With rapid development of communication technique, the communication system themselves and the modulation types they adopt are of more and more complexity and variety. All the factors above result in that the routine theory and methods of recognition can hardly meet the needs of practical signal recognition, thus the requirements to the methods and algorithms become higher and higher.Clustering analysis can discover the distribution characteristics of data and valuable relationship between the data attributes. Artificial Neural Network (ANN) has the abilities of instinctive nonlinear mapping based on standard data and powerful pattern recognition. Therefore, a deep study on these two artificial intelligence technologies is carried out in this paper including the design and implementation of an automatic modulation recognition system.The main work and contributes in this thesis can be summarized as follows:1. A novel modulation recognition method is proposed, which takes the constellation shape as classification characteristic. A new clustering method, named EAFCM, based on Fuzzy c-means algorithm is used for rebuilding the constellation. It not only solves the problem of sensitivity to initial centers as FCM, but also fixes the correct centers adaptively. Applying this method to modulation recognition of PSK/QAM signals has been shown by experiments to be feasible.2. Based on a study of the structure and learning algorithm of RBF and SOFM Neural network, an optimal learning algorithm of RBF network is proposed, and also, a combination model of SOFM network which takes advantage of LVQ algorithm is discussed: Application of both these two methods to modulation recognition of practical signals has shown improved ability of generalization. Compared with improved RBF neural network, the algorithm based on combination model of SOFM and LVQ possesses better performance of classification, and the average rate of correct recognition of each signal reaches 90%.3. Feature selection and Neural Network Ensemble are Studied and discussed in detail. Features selection are implemented by the use of Adaptive Genetic Algorithms, and based on the selected optimum and sub-optimum solutions (feature sets) a new method of Neural Network Ensemble named AGANNE is presented. The objective of AGANNE is to minimize the generalization errors of neural network and to increase the ensemble diversity simultaneously. Experimental results show that the feature selection can improve the classification performance of neural network, on the other hand, AGANNE achieves better accuracy under various SNR and dramatically improves the generalization ability of classifier, especially under moderate SNR, the average rate of correct classification reaches 94%, thus can meet the requirements in practical engineering.