Application Research Of Deep Learning In Radar Signal Processing

The increasingly complex electromagnetic environment of the battlefield and unknown target characteristics have brought more challenges to radar signal processing.Traditional radar signal processing is model-based,which mainly uses the prior information of the model and related signal processing criteria to design the signal processing method.It uses Gaussian,linear and stationary assumptions.Deep learning is a data-based method,which can directly obtain the relationship between input and output.This paper studies how to use this method in radar signal processing.The main research contents of this paper are: the neural network detector of deterministic signals under Gaussian and non-Gaussian noise conditions,the neural network detector for moving targets,and the detection performance of the neural network detector used in the measured data of early warning radar.This paper first regards the binary detection problem as a binary classification problem,and establishes a detection framework for signal detection using neural networks.A loss function including penalty terms is proposed to make the detector meet the Neyman Pearson criterion used in radar target detection,and the training strategy and network structure of the detector are given.Using Monte Carlo experimental methods,the detection performance and false alarm rate of the detector in the simulated data and the measured data of the early warning radar are estimated,and the influence of the measured data on the traditional detector and the neural network detector is analyzed.Compared with the classic methods,there are mainly matched filtering,optimal detector and DFT-based pulse Doppler processing.This article mainly has the following conclusions:1.By adjusting the weights of the loss function proposed in this paper,neural network detectors with different false alarm rates can be obtained when the network training converges.2.For signal models with optimal detection,such as deterministic signal detection under Gaussian white noise,neural network detectors can approximate the optimal detection performance.For models that cannot be optimally detected,such as the detection of nonGaussian and moving targets,the performance of this method is superior to existing methods.3.For the signal detection of uniformly moving targets,the detection performance of the neural network detector is superior to the DFT-based pulse Doppler processing method.4.The PDF and quantization accuracy of the measured data of the early warning radar will affect the detection performance.Although the noise in the measured data does not follow the Gaussian distribution,the performance of the neural network trained by the measured data is basically consistent with matched filtering.When AD is performed on the measured data and the quantization accuracy is too low,the performance of the detector will decrease significantly.