Radar Clutter Simulation Based On Deep Generative Learning

With the improvement of radar resolution and the diversification of radar working environment,radar clutter shows non-Gaussian,non-stationary and other characteristics.Traditional clutter simulation methods are difficult to simulate these characteristics comprehensively and accurately.In response to the above problems,researchers use the powerful learning and generation capabilities of deep neural networks to learn various characteristics of clutter,thereby simulating and generating more realistic clutter data.This method has gradually become a research hotspot.In this thesis,Deep Convolutional Generative Adversarial Network（DCGAN）is used for clutter simulation task,the model of clutter generation is constructed,clutter generation and conditional control are realized,and finally high-precision simulated clutter is obtained,which verifies the great potential of DCGAN in the field of clutter simulation.The main research content of the full thesis is summarised as follows.1.The characteristics of clutter are studied.The statistical characteristics of the measured sea clutter are analyzed based on relevant domestic and foreign sea clutter measurement experiments.To address the problem that the measured clutter data are not in a uniform format and cannot be used for DCGAN training,the measured clutter data is split and normalized to make the measured clutter data set in a standardized format while retaining the non-Gaussian and non-stationary characteristics of the original clutter data.2.The clutter simulation method based on statistical characteristics is studied.The working principle and process of the traditional clutter simulation method are sorted out,and the measured clutter data is fitted,modeled and simulated by the traditional method.The experimental results show that the K distribution model can better fit the amplitude distribution of the measured clutter data,and the simulated clutter data are obtained by fitting the K distribution,and its statistical characteristics have good fidelity.3.The DCGAN model is used to simulate the clutter with high accuracy to address the problem of accuracy loss of clutter obtained by traditional methods.The residual structure is integrated into the DCGAN model,which improves the stability of the DCGAN model.The upsampling module is used to replace the transposed convolution which solves the inherent "chessboard effect" of DCGAN and improves the randomness of generating two-dimensional clutter.Compared with the two benchmark methods,the traditional clutter simulation method and the Wave GAN generation model,the clutter generated by the DCGAN model is more realistic,and has better performance in amplitude distribution characteristics,correlation,non-stationarity,etc.It is also better than the two benchmark methods in the quantitative MMD comparison.4.The clutter generation model under continuous conditions is investigated using a simulated clutter dataset.The conditional control module is added to the DCGAN model to realize the clutter generation under continuous conditions.The difficulties of the clutter generation task based on regression labels are analyzed,and the solutions are given.Two label embedding methods,Reg-DCGAN and fast-Reg-DCGAN,are proposed to complete the clutter generation task based on regression labels,and both models perform much better than the traditional CGAN model in terms of label consistency.In addition,the label embedding method of fast-Reg-DCGAN is non-intrusive,suitable for any GAN architecture,and the training speed is also 16% faster than Reg-DCGAN.