Low Probability Of Interception Radar Modulation Recognition At Low SNR

Low probability of interception radar can control the transmission power to prevent the radar being intercepted and recognized.Waveform recognition of such signals is very important in electronic warfare.In the thesis,the waveform recognition algorithm is stud-ied under the condition of low SNR.Combined with time-frequency analysis and convo-lutional neural network,the recognition accuracy is further improved,and verified on the simulation data set of 10 modulations.At the same time,the convolutional neural network is understood from the perspective of signal analysis,which improves the interpretability of the algorithm.The main research contents and contributions of this thesis are as follows:1.To solve the problem of low recognition accuracy at low SNR,a recognition al-gorithm combining WVD transform and neural network based on MobileNet1.0-V2 was proposed.Under the verification of 5 random repeated samples,compared with the previ-ous research results,when the SNR is -8db,the method in this thesis is about 20% higher than the method based on LeNet,at -10dB about 3.5% higher than the method based on ResNet,and maintains a good result in time and memory consumption.2.There are many kinds of time-frequency analysis methods,which are difficult to choose.By presenting and comparing the commonly used time-frequency analysis meth-ods,it is found that Wigner-Ville transformation is more suitable,as the input of convolu-tion neural network with a strong recognition ability.And in the future experiments,it is preliminary confirmed.3.It is difficult to choose the convolutional neural network due to its variety.By presenting and comparing the five classical networks,it is found that the assessment of accuracy,time consumption,and memory consumption on ImageNet have a certain posi-tive correlation,with the corresponding assessment on the low probability of interception radar signals.It has the meaning of potential guiding the selection networks.4.In view of the lack of interpretation of convolutional neural network,it is re-described from the perspective of signal analysis.Subsequently the convolutional kernel,pooling,their parameters,the cascade and decomposition between convolutional nuclei,and the comprehension of convolutional neural network on wigner-ville transformation are expounded respectively.It improves the interpretability of the convolutional neural network algorithm and has the potential significance of guiding the design and use of convolutional neural network.5.On the basis of the redescription of the convolutional neural network,the time frequency uncertainty theory in the field of signal analysis is qualitatively extended to the convolution kernel in the convolutional neural network,obtaining the uncertainty of the locality of features(the size of the convolution kernel)and the fineness of features(the texture of the convolution kernel).It has the potential guide to the design and use of networks.