Research On Multi-Feature-based Radar Signal Classification

Radar emitter signal classification is a key process in electronic warfare, and it's also a hot and difficult issue in signal processing. In the modern electronic warfare, a variety of new and complex radar system into use, making the electromagnetic environment becomes more complex, but also bring more challenges to the radar emitter signal classification and recognition.On the basis of modern radar signal classification analysis, this paper discusses the general process of classification and recognition as feature extraction, feature selection and classification basis. For 4 common radar signals, the paper respectively extracts and analyses a variety of features from the time domain, frequency domain, bispectrum domain and wavelet domain, simulation experiments give the relation curves about the feature parameters at different signal to noise ratio(SNR), according to this, we can qualitatively determine the possibility of distinguishing different features. In order to improve the recognition efficiency, this paper bases on principle component analysis, gray correlation and uncertainty methods to screen features, and then to select the characteristic factors which contain the most classified information, finally using SVM method to classify the selected feature vector, the result shows that this classification method has a good classification results on the radar signals, in the meanwhile feature selection methods can shorten the running time of the classifier in ensuring the recognition rate, it also improves the classification efficiency.Furthermore, ECM method which is often used in image processing is introduced into this paper, to determine the different features of membership probability and fuzzy probability for each signal. Simulation results shows that, ECM method can accurately judge the membership probability of different features, according to the membership probability,we can realize the purposes of signal classification. To further verify the ECM classification results under different SNRs, the simulation shows the frequency domain features of skewness membership probability and fuzzy probability at two different SNRs. At last from the perspective of signal time-frequency distribution, this paper compares short-time Fourier transformation, WVD transformation and SM transformation, according to the principle of resolution and cross-term, a means is selected which is based on SM transformation to detect and estimate multi-source signals overlapped both in time domain and frequency domain. Simulation results show that the detection method which is based on the SM frequency transform can accurately estimate components signals.