Research On Radar Signal Recognition Algorithm Based On Lightweight AlexNet Model

The complex electromagnetic environment of modern battlefield and the emergence of new radar systems have led to the deterioration of the working environment of electronic reconnaissance and the rapid increase of the modulation mode and complexity of radar modulation signals.However,the traditional identification technology based on conventional signal parameters has been unable to meet the needs of various tactical activities in modern battlefield.Therefore,by referring to the success of in-depth learning in image recognition,this paper proposes a radar modulation signal recognition method based on time-frequency analysis and parameter compression lightweight AlexNet model under low signal-to-noise ratio(-12dB~0dB).Firstly,by analyzing and comparing the advantages and disadvantages of the common time-frequency analysis methods such as short-time Fourier transform,wavelet transform and smoothing pseudo-Wigner-Ville transform,the smoothing pseudo-Wigner-Ville transform is selected to analyze the conventional signal(NS),two-phase coded signal(BPSK),linear frequency modulation signal(LFM),non-linear frequency modulation signal(NLFM),frequency hopping signal(Costas),and multiphase coding.Ten kinds of radar signals(Frank)and multi-time code signals(T1-T4)are analyzed in time-frequency domain to generate time-frequency images of radar modulated signals.Next,adaptive wavelet threshold shrinkage denoising is applied to the generated time-frequency image to remove the noise in the signal.Then,the nearest neighbor interpolation,bilinear interpolation and bicubic interpolation are analyzed and compared,and the bicubic interpolation algorithm is selected to clip the denoised time-frequency image,thus reducing the redundancy of the image.Then,the structure layer parameters and the number of parameters of AlexNet,Google Net,VGG16 and ResNet,which have won the first and second place in the ImageNet image classification competition in recent years,are analyzed and compared.AlexNet model is selected as the algorithm model and the structure and parameters of AlexNet model are compressed and optimized,so that the time-frequency image pre-training of radar modulation signal can be accelerated.The simulation results show that the overall average recognition rate of 10 radar modulation signals is 89% when the signal-to-noise ratio is-12 dB,using the lightweight AlexNet algorithm model and through grid search test.Then,through changing the parameters and structure of AlexNet model,the lightweight AlexNet model is trained for the time-frequency images of 10 kinds of modulated signals after clipping.Then the parameters are selected by grid search to make the training model with the best cross-validation recognition rate.Finally,the time-frequency images are classified and recognized.The simulation results show that the method can automatically extract and select the features of time-frequency images at low SNR,and has a good correct recognition rate for up to 10 radar modulation types.The overall recognition rate of signal-to-noise ratio at-12 dB is 89%.Finally,aiming at the problem that the softmax classifier of the lightweight AlexNet model has insufficient generalization ability under low SNR and the training time of the model is still too long,this paper improves the lightweight AlexNet model by replacing the softmax of the lightweight AlexNet model with the extreme random forest classifier ERF.The simulation results show that the improved classification model has better generalization ability than the non-improved classification model,and the correct recognition rate is higher under low SNR.The overall recognition rate of signal-to-noise ratio is 91.55% and 2.55% higher under-12 dB,which further verifies the validity of the method.