Research On Application Of Neural Network In Recognition Of Radar Emitter Signal

With the rapid development of electronic information technology and electronic warfare,radar warfare system has become the research hotspot of current informational and intelligent electronic warfare.One of the most important functions of the radar warfare system is the identification of the radar radiation source signal,which analyzes and identifies the intercepted radar signal to judge the type,purpose,carrier,danger level and identification accuracy of the radar.Therefore,efficient radar radiation source signal identification is essential for subsequent radar analysis and operational preparation.With the in-depth study of radar system,the modulation methods of radar signal are increasingly diverse and the electromagnetic environment is more and more complex.Therefore,it is important to find an effective method to identify the radar radiation source in strong noise environment.When facing the recognition of radar radiation sources under strong noise,traditional methods often face the drawbacks of large amount of calculation of manual extracting features,strong subjectivity,invalidation under low signal-to-noise ratio and requiring researchers to have a large amount of knowledge about radar signal processing,etc.The performance of the traditional methods is difficult to meet the requirements.With the rapid development of neural network in the past decade,it has been widely used in signal processing and recognition.Therefore,this paper takes radar radiation source identification based on neural network as the research object,and conducts specific research.The main research contents are as follows:(1)To overcome the difficulty of complex radar radiation source signal recognition under low signal-to-noise ratio(SNR),a new radar signal recognition method based on improved deep residual network is proposed.First,the radar signal is converted into a two-dimensional timefrequency image by the Choi-Williams time-frequency transform to reflect the feature of the signal.Then,time-frequency image preprocessing and denoising convolution neural network are used to denoise the time-frequency image.The denoised image is input to the improved coordinated attention residual network for feature extraction,and finally the Softmax function is used to realize the identification of radar radiation source signals.In this paper,a timefrequency image dataset of nine of radar signals is established and compared with other algorithms.The experimental results show that the algorithm is insensitive to noise and has a good recognition rate at low signal-to-noise ratio.(2)Aiming at the problem that it takes a long time for radar signal time-frequency transformation to obtain time-frequency image,a direct use of two channels radar IQ time domain signals as the input of the algorithm is proposed to enhance the real-time performance of the algorithm.Soft threshold denoising is implemented in the deep residual network,and a small neural network is used to automatically calculate the threshold value,which adaptively eliminates the noise information in the feature learning process.At the same time,in order to better utilize the spatiotemporal correlation in radar signals,a bi-directional long short-term memory network and a self-attention mechanism are introduced to improve the performance of the proposed residual neural network.Nine kinds of time domain signal datasets of radar modulated signals are established.Experiments show that the algorithm performs well at low signal-to-noise ratio,which provides a new idea for radar radiation source identification at low signal-to-noise ratio.(3)To solve the computational complexity of the above algorithms,a lightweight model recognition algorithm based on knowledge distillation is proposed to increase the applicability of the algorithm on some devices with insufficient computing power.In the process of knowledge distillation,not only the soft labels predicted by the teacher network are used,but also distills the middle features of the teachers’ network.The algorithm proposed above is used as the teacher network,and the lightweight models ShuffleNet and MobileNet are selected as the student network.The experimental results show that the lightweight model after knowledge distillation performs better than the direct training lightweight model and reduces the model complexity.