Research On Radar Phase Error Correction And Imaging Technology Based On Deep Learning

In recent years,the rapid development of deep learning technology has provided new information processing methods and ideas for radar imaging.Unlike traditional physical model-driven imaging methods,data-driven deep learning techniques autonomously extract high-dimensional features from data samples to establish implicit mapping,which eliminates the dependence on explicit physical models.Therefore,deep learning has stronger expressive power and can represent more complex nonlinear mapping relationships,providing new advantages and possibilities for radar imaging in complex electromagnetic environments.However,the existing deep learning-based radar imaging networks overly dependent on training data samples,which limits their practical applications.Thus,it is necessary to develop new deep learning radar imaging frameworks with few or even zero samples.In addition,the highly nonlinear network models have poor interpretability,and the large number of learning parameters mean huge computational costs.This work focuses on the preliminary exploration of two themes:radar imaging in complex electromagnetic environments and model-driven deep learning radar imaging,aiming to improve imaging spatial resolution and reduce or even eliminate the network’s dependence on training data samples,and the radar imaging in ionospheric and atmospheric turbulence environments and model-driven deep learning are mainly studied.Firstly,the effect of ionospheric on low-frequency radio wave transmission is analyzed,and the relationship between the contaminated phase introduced by ionospheric effect and the frequency of electromagnetic waves and the total electron content(Total Electron Content,TEC)is derived.In order to solve the problem that the low efficiency and poor adaptability to ionospheric spatiotemporal variations of traditional ionospheric correction methods that estimate first and then correct.A P-band ISAR imaging ionospheric correction method based on Unet is proposed.The polar formation algorithm(PFA)is used to preprocess the echoes to reduce the burden of network training,and Unet is used to reconstruct high-resolution target images from coarse imaging results.Simulation results show that the proposed method effectively improves the imaging range resolution and has strong adaptability to ionospheric spatiotemporal variations.Secondly,the vortex electromagnetic waves based on the modulation of orbital angular momentum(OAM)and the principle of electromagnetic vortex imaging is introduced.The simulation model for atmospheric turbulence transmission was established,and the effect of atmospheric turbulence on vortex electromagnetic wave transmission and imaging is analyzed.The fluctuation of the refractive index results in spiral phasefront distortion,OAM mode diffusion,position shift,defocusing,and scattering intensity error in imaging results.To address these issues,a high-resolution imaging method based on Unet for electromagnetic vortex phasefront correction and turbulence compensation is proposed.Simulation results demonstrate that the proposed method achieves high-resolution focused imaging result and has better adaptability to spatiotemporal turbulence varying.Thirdly,utilizing algorithm unrolling techniques,the traditional iterative algorithm is combined with deep learning to optimize a small number of learning parameters in a data-driven manner.This approach enables the achievement of high-performance interpretable networks with a small number of training samples.Based on the Bayesian inference framework,the generalized expectation consistent(GEC)phase recovery algorithm is derived to solve the Bayesian estimation for ISAR imaging.To address the low efficiency and slow convergence of the GEC algorithm,the unfolding network GEC-Net is proposed,and only the damping factor is learned using a data-driven,while the other parts are solved as a Bayesian estimation problem.For the problem that the learning parameters of the unfolding network have layer-fixed limitations and poor adaptability to scene changes,a hypernetwork and attention mechanism are introduced to improve this problem.Depending on different inputs,the hypernetwork dynamically generates damping factors for the GEC-Net.Simulation results and measured data show that the method is highly flexible and can accurately reconstruct the target even in the test and training mismatch,avoiding frequent retraining.Finally,the expression of the point spread function(PSF)for ISAR imaging system is derived.For radar imaging applications in deep learning,it is often impossible to obtain the groundtruth of the target for supervised training.By using the physical model to guide network learning,the dependence on training data samples can be eliminated.The self-supervised learning ISAR high-resolution imaging method is proposed to incorporate the PSF of the ISAR imaging model into the loss function of the network,the method converts the direct image solution into an indirect solution of the minimized image generator network,and updates the network weights based on the measurement itself.Simulation results show that the proposed method achieves self-supervised high-resolution reconstruction for a single input.Moreover,the imaging model-guided loss function enables the iterative optimization process to converge quickly,improving computational efficiency.