Study Of Machine Learning In Fingerprint Identification Of Emitter Signals

Under the background of complex electromagnetic environment,the individual identification of emitter signals,as an important part in the field of electronic reconnaissance,has drawn more and more attention from domestic and foreign scholars.The individual identification of emitter signals,which is also called the "fingerprint" identification of emitter signals refers to the classification and identification of the specific emitter signals with the same intentional modulation parameters.However,at present,the specific identification rate of emitter signals is low and the utility is poor under the traditional methods of classification and recognition.Aiming at this,the thesis proposed a method of "fingerprint" recognition of emitter signals based on an integrated learning framework in machine learning.This dissertation started with extracting the "fingerprint" characterizations that characterize emitter signals.Based on the analyses of a single decision tree model,we studied two kinds of ensemble learning methods on the basis of the decision tree model and achieved the effective classification and identification of emitter signals under the ensemble learing framework.The main contents of this thesis include:1.According to the analyses of emitter signals characteristics,the "fingerprint" feature extraction of emitter signals was achieved through the transform domain methods,which specifically include the bispectrum "fingerprint" feature extraction based on higher-order cumulant and the "fingerprint" feature extraction based on reconstructed phase space.2.Based on the relevant theory of decision tree,the parameters were designed and an appropriate decision tree model was constructed.Then,multiple measure indexes for evaluating the classification performance of the decision tree classifier were given.After that,the classification and recognition of the extracted "fingerprint" characteristics of emitter signals was made by using the constracted decision tree model.3.As a single classification decision model,decision tree classifier easily leads to the problems of complex classification rules,convergence to non-global local optimal solution,over-fitting and so on.Aiming at this,a decision tree integration model based on Bagging method,that is,random forest was proposed.According to the recognition results of the two sets of measured data,it's found that the "fingerprint" identification rate of the emitter signals under the random forest model could reach over 96%.Compared with the single decision tree model,the random forest model exhibited better "fingerprint" recognition performance.4.In order to further improve the "fingerprint" recognition rate of emitter signals,another decision tree integration model based on the Gradient Boosting method,which is called GBDT was proposed.This model is generated through iteration of the CART tree,providing a better intergration framework of decision trees.From the recognition results of the two sets of measured data,it's found that the "fingerprint" identification rate of the emitter signals under the GBDT model could reach over 97%.In particular,99.7% or more of the "fingerprint" recognition rate could be achieved for the individual characteristics of the reconstructed phase space.Furthermore,we compared and analyzed the recognition performance and complexity of the three classification models of single CART decision tree,random forest and GBDT.According to the analyses,it was found that the "fingerprint" identification accuracy of the ensemble learning model was higher and the generalization performance was better under the same conditions.