Research On UAV Target Recognition Of Micro-Doppler Spectrum Based On Deep Learning

UAVs have been widely used in various fields due to their small size,easy access,and easy concealment.However,there are many potential risk in their applications which may endanger public security,such as the problem of black flying.Therefore,it is particularly important to identify UAVs,determine their potential hazards,and make timely responses.The rapid growth of deep learning in recent years and its trait of automatic feature extraction have provided possibilities for UAV target recognition.Thus,this thesis has conducted research on the method of UAV target recognition of MicroDooppler Spectrum based on deep learning.The main content is as follows:(1)The target echo model is established and then construct its Micro-Doppler spectrum database.Firstly,construct radar echo models for UAVs and birds by analyzing the micro-motion characteristics of UAV blades and bird wings.And then use STFT timefrequency analysis to obtain the micro Doppler spectra of targets’ echo signals.Finally,a Micro-Doppler spectrum database was constructed for five types of targets,namely,single rotor drone,four rotor drone,six rotor drone,eight rotor drone,and birds.(2)A method of UAV target recognition based on Feature Focus Convolutional Neural Network is proposed.According to the characteristics of the target spectrum,this method designs an attention subnet to force the network to focus on useful areas like the human eye.And then,embedds the attention subnet into CNN to build a Feature Focus Convolutional neural network named FFCNN to achieve UAV target recognition.Due to the embedding of the attention subnet,this network can focus on the regions in the target that are conducive to classification,extract more discriminative features,and thus improve the accuracy.Experimental results indicate that the recognition accuracy of FFCNN at a signal-to-noise ratio of 0d B is 1.67% higher than that of the network without embedded attention subnet.(3)A method of UAV target recognition based on Depth Threshold Residual Network is proposed.This method proposes a threshold residual subnet by combining the residual structure,and then constructs a Deep Threshold Residual Network named DTRSN by cascading multiple threshold residual subnets to achieve UAV target recognition.Because the threshold residual subnet can automatically learn to suppress interference features to a certain extent,this method can improve recognition performance at low SNR.The experimental results indicate that this method’s accuracy of recognition is still over 95% when the signal to noise ratio is-10 d B.(4)Studied the UAV target recognition under the small sample condition.By utilizing the idea of siamese network and using the backbone network of FFCNN as a feature extraction subnet,a network named Siam FF is constructed to achieve UAV target recognition under small samples.This network effectively avoids overfitting under small samples by continuously reducing the distance of similar sample pairs in the feature domain and increasing the distance of different sample pairs in the feature domain,thereby improving the accuracy of recognition under small samples.Experimental results indicate that the accuracy of Siam FF is 98.22%,which is about 10% higher than that of FFCNN network when the signal-to-noise ratio is 0 d B and the training sample size for each type of target is 60.