Research On Radar Signal MTI Based On Deep Learning And Sparse Representation

As military modernization accelerates,advanced digital waveform generation and processing techniques continue to emerge,allowing a large number of new radar waveforms to be designed and radar signals to become more diverse and complex.At the same time,the modern electromagnetic environment is becoming more and more complex,making radar waveform identification under extreme conditions increasingly important.Traditional modulation type identification techniques are quite complex and technically demanding,not very adaptable,and have been difficult to apply to the rapidly changing modern scenarios.Based on the above research background,this paper systematically conducts a study on radar signal modulation type identification based on deep learning and sparse representation.The research in this paper consists of three parts.1.First of all,the signal feature method is studied,and time frequency simulation is performed on common radar signals.Simulation of the signal under Gaussian white noise and ricean multipath damage environment revealed severe damage to the Choi-Williams time-frequency representation and Hilbert spectrum of the signal at low SNR.A dataset consisting of 11 radar signals was generated,then preprocessed time-frequency features used in this paper.2.Secondly,how to accurately denoise from noised radar waveforms was studied.A Denoised ODL algorithm(DNODL)is proposed based on the signal sparse a priori.The simulation results show that DNODL has a better performance than the ODL algorithm.The reconstruction errors of the DNODL algorithm for different dictionary sizes and different sparsities are investigated,and the signal sparse denoising is performed at optimal sparsity and dictionary size.The simulation results show that the sparse denoising method proposed in this paper improves the average SNR of the signal from-10dB to-0.83dB.Due to the structure retention feature of the learning dictionary,the reconstructed signal structure information is complete.A Discriminative ODL algorithm(DODL)is then proposed based on online learning ideas,where a linear classifier is fused for classification judgments based on sparse representation coefficients.DNODL and DODL fusion identification is used.The results showed that the linear classification method was able to achieve 64.5%accuracy at-10dB.3.Finally,deep learning methods are introduced.A modulation classification algorithm based on time feature(TMC)is proposed,and the simulation results show that 95%accuracy is achieved at SNR 0dB and above,but the accuracy decays rapidly with the SNR decreases.A single Choi-Williams time-frequency representation classification method was simulated and found to be inadequate for the separation of similarly modulated signals.A modulation classification algorithm based on joint time-frequency feature constituted with Choi-Williams time-frequency representation and Hilbert spectrum(JTMC)is proposed.The simulation results show that 81.7%accuracy can be achieved at-10dB;the SNR of the input signal is further improved by DNODL sparse denoised algorithm,and a modulation classification algorithm based on sparse denoising and joint time-frequency feature(DJTMC)is proposed.The algorithm achieves an accuracy of 92.1%at-10dB.The final simulation comparison verifies the validity of the method presented in this paper.