Performance Analysis Of Target Localization For Distributed MIMO Radar Based On Supervised Learning

The transmitting and receiving antennas of MIMO(Multiple-Input Multiple-Output)radar with widely separated antennas are far apart in space.MIMO radar possesses waveform diversity gain and spatial diversity gain which are not available in traditional radar,which can significantly improve the system performance of target detection and parameter estimation.Therefore,it has attracted extensive attention and research.The traditional algorithm for target parameter estimation has high computational complexity which is not controllable.In the complex and changeable actual environment,the performance of adaptive parameter estimation based on real-time received data is also limited.Deep learning has a powerful ability of real-time data processing and learning,which has certain advantages over traditional methods in some cases.To facilitate analysis without loss of generality,typical fully connected neural network is used to estimate target position parameters in this paper.However,the estimation performance of different topological structures is different,and the selection of network topological structures only depends on experience,which cannot provide the prediction of the performance or the theoretical support.To solve this problem,this paper first studies the performance bound of fully connected neural network based on the prior statistical characteristics of sample set for the regression task,then analyzes the performance of the target localization of distributed MIMO radar based on traditional statistical methods,and finally studies the performance of the target localization of distributed MIMO radar based on fully connected neural network.(1)First for regression tasks,this paper introduces the topology structure of the fully connected neural network,then assumes the prior statistical characteristic of the sample set.When given the network topology,the initialization parameter and the statistical characteristics of the sample set,the training performance of the fully connected neural network is studied.This paper respectively analyzes the forward propagation process based on the ADF(Assumed Density Filtering)model and the prior statistical characteristic of sample set and the back propagation process based on Adam(Adaptive moments).In addition,the testing performance is studied.Finally,considering the statistical characteristic of the initialization parameters,the MSE(Mean Square Error)bound of the network performance is derived.Based on this MSE bound,the optimization of the neural network topology structure can be achieved.(2)Consider traditional statistical methods for target localization.The received signal of the MIMO radar with widely separated antennas is modeled firstly,then the distributed MIMO radar estimator for target position estimation is introduced,then the ML(Maximum Likelihood)estimation and CRB(Cramér–Rao Bound)of the time delay are derived respectively,finally the CRB of the target position is derived based on the distribution of the time delay estimates.The correctness of CRB is verified by simulation.(3)Consider deep learning methods for target localization.The fully connected neural network is used as the core unit of the data fusion center.The statistical characteristic of sample set for target localization is analyzed and the network topology is designed.Then the process of target localization algorithm based on the statistical characteristic of sample set is studied.Finally,the MSE bound for target location estimation is derived and the network topology structure is optimized.The validity of the MSE bound of the neural network is verified by simulation,and the effects of network topology,location sampling interval and number of transceiver paths on the MSE bound are analyzed.And the MSE bound of the network is campared with the MSE bound of traditional statistical methods.