Research On Application And Parallel Implementation Of Data Mining In Modulation Type Recognition

Classification and clustering are two commonly used data mining algorithms which can find out the valuable information from a large number of data.Classification is a kind of supervised learning algorithms,it can classify the data sets according to the criteria which is learned from the training sets.Clustering is a kind of unsupervised learning algorithms,it can find the distribution structure of data sets and divide the data into the clusters without prior knowledge.This dissertation mainly studies modulation type recognition methods by using data mining algorithms and their parallel implements.Firstly,the research background,the key technology of data mining and modulation recognition are introduced.k-means algorithm,DBSCAN algorithm and RBM model are detailed.Secondly,the general formulas of the high-order moments of MASK,MPSK,2FSK,MQAM are derived.And two feature parameters are extracted from the high-order moments of 2ASK,4ASK,QPSK,8PSK,2FSK,16 QAM which are calculated by the general formulas.The digital modulation signals recognition method based on hybrid three-order restricted Boltzmann machine(H3RBM)is proposed.Simulation results show that the training time of H3 RBM is short and the recognition rates are high.Thirdly,a method of modulation type recognition based on parallel k-means algorithm is proposed.By random sampling,the data set is divided into multi blocks.The pre-clustering optimizes the quality of initial cluster centers,and the linear prediction of average cluster centers improves the clustering efficiency.The proposed method can identify 2ASK,4ASK,QPSK,8PSK,2FSK,16 QAM without training sets.Simulation results show that the proposed parallel implement of k-means algorithm has a higher clustering efficiency and the proposed method has high recognition rates.Finally,a method of modulation type recognition based on parallel DBSCAN algorithm is proposed.Particle swarm optimization is used to search the optimal boundaries based on the data overlap partitioning.The clustering of each data block is achieved by rough clustering and rough cluster merging to reduce computational complexity.The proposed method can identify 2ASK,4ASK,QPSK,8PSK,2FSK,16 QAM without training sets.Simulation results show that the parallel implement of DBSCAN algorithm effectively improves clustering efficiency and the proposed method has higher recognition rates than that of the recognition method based on k-means and the recognition method based on H3RBM.