Study On Automatic Modulation Recognition Algorithms Of Digital Modulated Signals

Digital modulation recognition is the key technology of the signals demodulation,information extraction and signals detection in communication systems, so it is veryimportant in both cooperation and non-cooperative communication fields.Based on the analysis and summary on the related research at home and abroad,feature extraction and classification are given further study so as to solve the importantissues in modulation recognition.The main contents of this thesis include:Firstly, two popular feature extraction methods based on instantaneous informationand high order cumulants were researched and realized. A group of excellent combinedfeature parameter sets were proposed based on the two methods. However, theinstantaneous information was easily affected by noise, so improved wavelet thresholdde-noising algorithm was used to optimize it.Secondly, the classifier’s performance based on decision tree was studied andrealized. According to the combined feature parameter sets, recognition flow of tendigital modulated signals based on decision tree classifier was designed, then theclassifier’s recognition performance was simulated and analyzed. The simulation resultsshow that the algorithm can get high recognition probability when the inputsignal-to-noise ratio(SNR) is5dB. However, decision tree classifier is too dependent oneach parameter and affected by the thresholds, so it cannot get good performance in lowSNR conditions.Thirdly, to solve the shortcomings of the decision tree classifier, neuralnetwork(NN) was introduced, the classification principles of back propagation(BP) NNwas analyzed. Three improved algorithm were adopted and simulated, then, resilientback-propagation(RPROP) was chosen as the training algorithm of the BP NN. Thecomputer simulations show that the NN classifier can obtain an overall success rate of98%at the SNR of-2dB, but there are obvious shortcomings, such as: less learning,over learning and a local minimum value, etc.Finally, in order to overcome NN classifier’s disadvantages, a support vectormachine(SVM) classifier was designed. After analyzing the basic theory andclassification principle, based on the same samples with the two above-mentionedclassifiers, the classification performance was simulated with different number of sample sets, different SNR and different sample length. The simulation results show thatthis classifier algorithm can get an overall success rate of86%when the SNR is-5dBand the data length is500. In addition, compared the NN classifier and SVM classifierunder the conditions of low SNR and small sample sets, SVM classifier can get betterClassification performance.