| With the complexity of modern warfare environment,and the diversification of hostile targets and tasks,how to accurately identify targets in harsh environments is a hot issue in the field of radar research.In the research of radar target recognition,highresolution one-dimensional profile(HRRP)has been widely concerned because it can better reflect the geometric structure information of radar targets,and has the characteristics of easy acquisition and strong real-time performance.This thesis mainly studies the target recognition of HRRP radar based on deep learning algorithm.The main research contents include:Firstly,a novel Concatenation Deep Neural Network(CDNN)model is designed by analyzing the deep learning algorithm and features of HRRPs.In the construction of the CDNN model,the depth of the network is deepened by cascading multiple sets of shallow sub-networks,so that the high-order abstract features of HRRP can be extracted;the subnetwork is initialized by means of network parameter migration to accelerate the training process of the CDNN model.In addition,according to the aspect sensitivity problem of HRRP,this thesis proposes a Secondary-Label(SL)coding scheme.By dividing the target’s aspect,the corresponding primary and secondary labels are set for the samples of each type of target.This coding method makes it not only consider the difference between the classes of the target in the process of training,but also considers the differences within the class.In the model testing phase,a decision fusion strategy is adopted to comprehensively identify continuous HRRP samples.Secondly,aiming at the small target HRRP radar target recognition problem,this thesis proposes a feature domain migration model DTNN by improving the migration network.In the DTNN model,two homomorphic sub-networks are included,and the recognition rate of the measured HRRP data is optimized by using the simulated HRRP data.During the training of the model,the two sub-networks are alternately trained and share parameters.At the same time,by setting the MMD distance,the difference between the feature distribution of the source domain and the target domain is minimized when the model recognition error is minimized.In this way,the feature domain migration between the source domain and the target domain is realized,so as to achieve a higher recognition rate with only a small amount of measured data.Finally,an improved generated confrontation network(GAN)is designed for data enhancement for the problem of imbalance of sample numbers between different targets in HRRP radar target recognition,that is,generates the required samples through WGAN and CWGAN.In the design of WGAN model,EM distance is introduced as the training index of the generated model to improve the quality of the sample generated by the model.In the CWGAN model,the label condition constraint is introduced to improve the efficiency of the WGAN model to generate samples.Through the study of deep learning,this thesis further broadens the application of neural network algorithm in HRRP radar target,which also provides a new idea for future radar identification research. |
