Multi-kernel And Multi-task Learning For Radar Target Recognition

Due to the capability to produce all-weather and 24-hour a day towards environmental condition,radar sensors play an important role in the modern battlefield.Radar automatic target recognition(RATR)technology is realized by extracting the relevant information characterizing the target of radar echo signals to determine the category of targets.RATR is one of the important sources of enemy intelligence and can provide strong decision-making information for our army during the battle.With the development of high-resolution radar systems,it becomes possible to obtain more accurate structures and detailed information of targets.However,high-resolution radar images have the problems of target attitude sensitivity,high sample dimensions and non-linearity of scattering structures.It is difficult for traditional algorithms to accurately recognize radar targets.In this paper,a related research on RATR is made based on the theory of multi-kernel learning and multi-task learning.In the modern warfare,the electromagnetic environment has become more and more complicated.Single-functional electronic equipment can no longer meet the needs of combat missions.As two major electronic equipments,radar and communication systems are getting closer in system architecture,frequency bands,and bandwidth.The integration of radar and communication has become a trend.In this paper,communication data is added to radar waveform to form an integrated waveform and the RATR performance of the integrated waveform is preliminarily studied.The main research contents of this paper include:1.A multi-kernel learning(MKL)and multi-task learning(MTL)for radar target recognition is proposed.For MKL learning,the data-dependent kernel function is adopted as the basic kernel function.Then the criterion of kernel alignment(KA)is used to construct an adaptive weighted multi-kernel function.Compared with conventional kernel functions,the adaptive kernel function has a better interpretation capability of data.In this paper,the recognition of each target is considered as a task.By adopting the mixed norms lp,q,MTL is realized and information among multiple tasks can be shared.In the mixed norm,the sparsity of the characterization of the kernel functions among different tasks is controlled by lp,and the tightness of the connections among different tasks is constrained by lq.Compared with the fixed norm l1,2,the mixed norm lp,q has a better adaptability.The combination of MKL and MTL can effectively increase the flexibility of model.To find the solution of model,genetic algorithm(GA)is adopted to optimize the parameters of MKL and MTL model.2.A method based on orthogonal neighborhood preserving projection(ONPP)and maximum margin criterion(MMC)is proposed for radar target recognition.The dimension of radar image is high,and a large-scale training sample will impose a heavy burden on the classifier.Therefore,it is necessary to reduce the dimension of training data.Taking into account the highly nonlinear characteristics of radar targets,ONPP is adopted to reduce the dimensions,and then it is extended to the multi-kernel function field.Namely,ONPP is used to supervise the multi-kernel learning.ONPP can effectively reduce the sample dimensionality and maximally preserve the structure information,but without the discrimination.MMC can maximize the scatter of data from different classes and minimizing the scatter of data from the same class,which increases the separability of arget.Therefore,the joint of ONPP and MMC can not only effectively reduce the dimension,but also have a better recognition ability.3.A clustered multi-task learning for radar target recognition is proposed.Conventional multi-task learning methods assume that the connections among different tasks are the same,but in fact some tasks are more closely linked,and the others are not.In this paper,the relationship among multiple tasks are learned adaptively.According to the automatically learned relationships,multiple tasks are divided into several groups.The tasks are more closely related within the same group and weaker among different groups.Considering the highly nonlinear characteristics of radar targets,multi-task learning is extended to nonlinear fields.Inspired by the idea of ONPP,this paper assumes that the closely-related tasks in the original space still maintain the close relationships in the non-linear space,which can fully utilize the multi-task relations in the nonlinear space.To solve the MKL model,this paper proposes a solving method which can achieve parallel computing.4.An integrated waveform for radar communication recognition is designed.The waveform is realized by adopting orthogonal frequency division multiplexing(OFDM)and applied to RATR.Specifically,communication data is embedded into the sub-carriers of OFDM radar signal to form an integrated waveform.In terms of high peak-to-mean envelope power ratio(PMEPR),Gray code is adopted to reduce the PMEPR of the integrated waveform.However,the communication data and low PMERP in the integrated waveform will lead to high PMEPR and the reduction of the recognition performance of waveform.The cyclic shift sequence can effectively reduce the PSLR of the integrated waveform without affecting its PMEPR.Therefore,a combination of gray coding and cyclic shift can simultaneously realize a lower PMEPR and PSLR of integrated waveform.