Radar Emitter Signal Recognition Based On Time Frequency Analysis

The recognition of radar emitter signals is a critical process for electronic intelligence(ELINT) and electronic support system(ESM), which plays an important role in the entireelectronic warfare. The feature extraction technique is the precondition and foundation ofrecognizing radar emitter signals. Traditional methods of recognizing radar signals are generallybased on five conventional parameters including radio frequency(RF), time of arrival(TOA),pulse width(PW), pulse amplitude(PA) and direction of arrival(DOA), which recognize radaremitter signals by matching feature and parameters in radar signals database. However, with therapid development of electronic technology and radar technology, the advanced radar play adominant role in modern warfare, and the circumstance of radar signal become increasingdenseness. As a result, the performances of these traditional methods descend rapidly, whichcannot meet the requirements of modern electronic warfare. Therefore, it is necessary to explorenew and valid features for improving the accurate recognition rate of radar signals.Time-frequency distribution transforms time domain signals into two-dimensionaltime-frequency domain signals，which depicts the distribution law of emitter signal energy intime-frequency plane. So it can provide the detailed information for the feature analysis of radarsignal. To correctly classify advanced radar emitter signals in low signal-to-noise rate(SNR), theintra-pulse feature is investigated by time-frequency analysis. Then the time-frequency feature isproposed, and the method of instantaneous frequency (IF) estimation is also studied. The mainwork and innovations obtained in this paper can be summarized as follows.1. A novel approach adopting Rényi entropy of time-frequency distribution for radar emittersignal recognition is proposed. Time-frequency distribution of radar signal is obtained by usingsmoothed pseudo Wigner-Ville time-frequency transform, and then the third-order, fifth-order,seventh-order, ninth-order and eleventh-order Rényi entropy of time-frequency distribution areapplied to construct a feature vector, which has low dimensions and large between-classdifference for radar signal recognition. Finally the support vector machine is applied to identifyeight radar emitter signals automatically. Simulation results show that the proposed approach canachieve satisfying accurate recognition over a wide range of SNR scenarios. Even for SNR=-3dB,the accurate recognition rate still achieves90.75%.2. A novel approach using image feature of time-frequency distribution for radar emittersignal recognition is proposed, which transforms the classification of radar signals into imageprocessing and image recognition. Time-frequency images of radar signals are obtained by usingChoi-Williams time-frequency transform, and then these images are transformed into grayscaleimages. The geometry features of time-frequency image are discussed due to time-frequencyimages of radar signals have great difference obviously. A series of image processing methodsare employed for time-frequency image enhancement and de-noising. In addition, the areas notcontaining signal components from the edges of the image are removed. Finally, the centralizemoments and pseudo-zernike moments are calculated as the feature for signal recognition, andthe support vector machine is applied to identify eight radar emitter signals automatically. Simulation results show that the proposed approach can achieve satisfying accurate recognitionwhen SNR varies in a large range. Even for SNR=-3dB, the proposed method which adoptspseudo-zernike moments works effectively as high as92%recognition rate. The validity of theapproach is demonstrated by experiments result.3. The texture features are investigated for distinguishing signals whose time-frequencyimages are approximate. A novel approach using image feature of time-frequency distributionbase on the local binary patterns is proposed，and the support vector machine is applied toidentify twelve radar emitter signals automatically. Simulation results show that the proposedapproach can achieve satisfying accurate recognition in low signal-to-noise(SNR) rate. Even forSNR=0dB, the proposed approach achieves overall correct classification rate of95.35%. Theapproximate LFM signals is also recognized effectively.4. To estimate the IF accurately in the condition of low SNR, a hybrid time-frequencydistribution method is proposed, which combines the Wigner-Ville distribution(WVD) andChoi-Williams distribution(CWD) based on the Hadamard product of time frequency matrices.Simultaneously, the probability that the time-frequency distribution takes a maximal valueoutside the auto-term position is considered as the guide line for selecting the window length oftime-frequency analysis, and then the initial IF is estimated. The intersection of the confidenceintervals algorithm is applied to detect the IF along with initial IF finally. The proposed approachis employed to estimate the IF of SFM, LFM and FSK signals, and compared with the method ofthe first moment of the time-frequency distribution and the peak of the WVD. The validity of theproposed method is demonstrated by experiments result.5. To estimate the IF of multi-component signals in the condition of the low SNR, a hybridIF estimation method which combines the time-frequency analysis with image processingtechnique is proposed. Firstly, the time-frequency peak filtering is applied to attenuating randomnoise for signals, and then the one-dimensional signal is transformed to the two-dimensionaltime-frequency image using modified B distribution. The areas not containing signal componentsfrom the edges of the time-frequency image are removed by detecting the starting and endingfrequencies, and the leavings grayscale images are converted into binary images. The thinningalgorithm of morphology can lead to the improvement of time-frequency resolution and the noisesuppression. In addition, the skeleton of time-frequency image is obtained via skeletonizingalgorithm. Finally, the method of labeling connected components is adopted for marking thedifferent signals and noise components. Further more, the IF is estimated by finding the row andcolumn indices of signal component. The proposed approach is employed to estimate the IF ofSFM, LFM and FSK signals. Simulation results show that the proposed approach is efficient inthe condition of low SNR and flexible in processing multi-component signals.