Research On Feature Extraction And Recognition Of Radar Signals

With the rapid development of radar technology and the improvement of anti-jamming capabilities,radar signal types have become increasingly complicated and varied,making it increasingly difficult for traditional electronic reconnaissance systems to recognize modulation of intercepted radar signal.Extracting the intra-pulse features of radar signals that are more effective and more universal as a supplement to the classic five-parameter is an effective way to improve radar signal sorting and identification.Therefore,for the intercepted unknown radar pulse signal,theoretical problems related to the feature extraction and identification of radar emitter signals are discussed from three aspects: ambiguity function,zero-Doppler cut(autocorrelation function)of the ambiguity function,and classifier selection.Main work and research results are as follows:1.Typical radar signal models are established,including conventional pulse,frequency modulation,and phase modulation signals.The time domain,frequency domain and the ambiguity function of the signals are simulated and analyzed.Besides,the classifier models for identifying waveforms are analyzed.2.The signal interception process is analyzed and the intercepted signal model is established.The parameters such as carrier frequency,amplitude,and initial phase of different intercepted radar signals may be different.In order to extract the features which are invariant to these parameters,a feature extraction method based on autocorrelation(ACF)is studied,and k-nearest neighbor(k NN)and support vector machine(SVM)classifiers are used to identify signal modulation type by the feature vectors.When classifying with k NN,the recognition rate can reach more than 80% when the signal to noise ratio(SNR)is-2d B.When processing the same data,SVM has better classification performance than k NN.In order to reduce noise interference,the ensemble average ACF of the intercepted pulse groups are computed.The simulation results show that this method can significantly improve the recognition rate of the signal.Both classifiers achieve a recognition rate of nearly 100% at 5 d B by averaging 20 ACFs together.3.Given that the auto-correlation function is equivalent to the zero-Doppler cut of the ambiguity function,and the autocorrelation does not include the information of the signal in frequency domain.In order to extract the invariant features to the foregoing parameters,the thesis substitutes the autocorrelation-based feature for the ambiguity-based feature with the same signal parameter settings.KNN and SVM classifiers are used to classify feature vectors based on ambiguity functions,and similarly achieve ideal classification results.Especially when using the SVM classifier,when the number of accumulated pulses can reach 1,5,and 10,the recognition rate is improved at low SNR compared to the feature extracted based on autocorrelation,which is particularly beneficial for signal recognition where pulse accumulation is difficult.