Research On Key Technologies Of Intelligent Recognition For LPI Radar Signals

In modern warfare,radar signal recognition is an important component of electronic reconnaissance systems and has received extensive attention.Conventional methods for radar signal recognition have a low recognition rate and are sensitive to varying signal-to-noise ratios(SNRs).Especially,with the continuous improvement of the electronic level of weapons and equipment,new system radars represented by low probability of intercept(LPI)radar are widely used,which makes the disadvantages of traditional radar signal recognition methods more obvious.In recent years,the rapid development of artificial intelligence technologies and various accelerated computing chips has provided new approaches for radar signal recognition.This paper mainly focuses on the key technologies of intelligent recognition of LPI radar signals.The specific research work is as follows:(1)Aiming at the problem that the conventional methods have a low recognition rate under low SNRs and to combine the advantages of neural networks,this paper proposes an intelligent recognition method for LPI radar signals with impact noise based on the convolutional neural network(CNN).The method firstly performs a timefrequency(TF)analysis on the radar signals to obtain TF images and then employs CNN to train TF images and predict the modulation mode of the radar signals.Experimental results show that the recognition accuracy of the CNN model proposed in this paper is slightly lower than that of the Google Net model,but its complexity is much smaller than that of the Google Net model,which is more conducive to the implementation on embedded devices.(2)Since the TF analysis is time-consuming due to a large amount of computation,this paper proposes a hardware acceleration scheme for the Choi-Williams distribution(CWD)time-frequency analysis algorithm based on Field Programmable Gate Array(FPGA).The scheme can greatly improve the computational speed of time-frequency analysis without significantly reducing the computational accuracy.Experimental results show that the hardware acceleration scheme can improve the computation speed by 49.29 times compared with the software computation.(3)Aiming at the high requirements of real-time,low-power,and miniaturization of airborne and missile-borne reconnaissance equipment,this paper uses FPGA to accelerate the calculation of compressed CNN,and a sparse matrix-based CNN hardware acceleration architecture is designed.Compared with the pulsed array-based FPGA convolutional neural network acceleration architecture,the new acceleration architecture proposed in this paper can provide a maximum acceleration performance improvement of about 4 times while using the same multiplier hardware resources.And the overall computing performance is 1.63 times that of the Core 7th generation i5 CPU,0.66 times that of the Core 10 th generation i7 CPU,and the energy efficiency is 7.88 times that of the Core 7th generation i5 CPU and 3.2 times that of the Core 10 th generation i7 CPU.