Research On Individual Identification Of Radar Radiation Source Based On Machine Learning

With the development of electronic technology,plenty of modern electronic equipment paly an import part in warfare,which makes the electromagnetic environment complicated.The individual identification of radar radiation source is essential for accurate counterattacks.However,the current radar radiation source identification method is parameter matching,which cannot meet the requirements both in accuracy and speed.This dissertation focuses on feature extraction and classification,the main contributions are described as follow:This dissertation proposes two radar radiation source feature extraction methods.Firstly,the transformation-based method calculates the statistical characteristics of the transformation of original signal,The other method is based on Autoencoder.We give the theoretical analysis of the proposed method and proving its better performance.The inputs of Autoencoder are signals which processed by Discrete Fourier Transform,and the outputs of hidden layers are used as features.Moreover,we adopt Population Based Training(PBT)to optimize the hyper-parameters.This dissertation proposes the estimation system of the radar radiation source features,which consists of complexity,separability and representational ability.The experiment demonstrates that the feature extraction method based on Autoencoder is less complicated while the method based on transformation is better in separability and representational ability.This dissertation proposes the radar radiation source classification method based on machine learning.Firstly,we analyze the performance of Extreme Gradient Boosting(XGBoost)and random forest in radar radiation source classification.Random forest is better both in speed and accuracy.Furthermore,the proposed Autoencoder-based feature extraction method can improve the precision.To explain the classification result,we apply SHapley Additive exPlanations(SHAP).Concerning the limit of time,this dissertation proposes the classification method based on SincNet.In SincNet,the first convolutional layer adds band-pass filters.SincNet converges faster than Convolutional Neural Networks(CNN).