Radar Emitter Identification Technology Based On Time-frequency Analysis And Deep Learning

Radar emitter identification(REI)is a key link to obtain intelligence information in electronic reconnaissance and occupies an important position in electronic warfare.With the development of radar technology,the conventional parameters of modern radar have the characteristics of complex,changeable,and overlapped seriously.Traditional parameter features-based REI methods are hard to identify the emitters accurately.Using intra-pulse features for recognition can solve the problem of complex,variable and overlapping conventional parameters.Most radar signals are non-stationary with low probabilities of interception and have complex waveforms.Based on these characteristics,the paper applies intra-pulse time-frequency(TF)features to characterize the signals.And aiming at the problem of TF distribution features are complex and difficult to be directly used,the deep learning network classifiers are used to identify the radar emitters.To promote the recognition performance under the low signal-to-noise ratio(SNR),the improved TF transforms are finally proposed and applied to REI.The main contents are as follows:1.TF analysis of radar emitter signals was carried out through short-time Fourier transform(STFT),Wigner-Villel distribution(WVD),smoothed pseudo Wigner-Ville distribution(SPWVD),and Choi-Williams distribution(CWD),and compared under different SNRs.The experiment results show SPWVD and CWD have the best TF analysis performance but the worst efficiency.WVD has the worst performance.STFT has the highest efficiency,and the average processing time of STFT is only 1.06%and 1.04%of SPWVD and CWD.2.The classifiers are constructed by using convolutional neural network(CNN)and gate recurrent unit(GRU)network.REI is realized by the input of TF spectrum,and the performance based on different TF distributions and based on CNN and GRU is compared.The experiment results show the SPWVD-based and the CWD-based recognition performance are the best.The average recognition rates of SPWVD and CWD are 1.26 and 1.25 times of WVD respectively.Among the two classifiers,CNN has better recognition performance,and the average recognition rate of CNN is 1.03 times of GRU.However,the training of GRU is more efficient,and the average training time of GRU is only 59.86%of CNN.3.Two improved TF transforms based on STFT are proposed and applied for REI:F-STFT and P-STFT.First,STFT is used to calculate the frequency spectrum of the signal.The frequency spectrum is combined with the k-medoids algorithm to filter the STFT distribution and the F-STFT distribution is obtained.Then,the auto-correlation function is applied to calculate the power spectrum of the signal.The power spectrum is combined with the k-medoids algorithm to filter the STFT distribution and the P-STFT distribution is obtained.The experiment results show F-STFT and P-STFT can promote the TF analysis performance and REI performance.F-STFT and P-STFT improve the average recognition rate of STFT from 0.83 to 0.94 and 0.96 under negative SNRs.F-STFT and P-STFT achieve higher recognition rates than SPWVD and CWD but still guarantee real-time performance.The average processing time of the two improved STFT is less than 5%of SPWVD and CWD.Among the two improved STFTs,the efficiency of F-STFT is slightly higher than P-STFT,and the recognition rate of P-STFT is slightly higher than F-STFT.