Research On Key Technologies Of Passive Localization Based On Deep Learning

Localization has been playing an important role in daily lives ever since ancient times,when people used stars and compasses as guidance for positions,evidenced in the sayings that "taking carriage loaded compass to avoid confusion","looking forward to the sun,moon,and stars".Nowadays,with the rapid development of wireless communication technology,passive positioning based on received radio signals is becoming increasingly popular.Among them,multi-station passive localization has been widely studied due to its short positioning time,high accuracy,and good stability.However,the multi-station passive localization faces critical challenges in its practical applications due to the constraints of cost,computing power,energy consumption,and transmission bandwidth.This thesis studies two key technologies to address these challenges including heterogeneous fusion localization and node selection,whereby the main research contents and contributions can be summarized in the following:Firstly,to address the problem that the obtained measurements are heterogeneous due to practical constraints such as thepower consumption and computational complexity,a multi-station positioning algorithm which can flexibly fuse heterogeneous parameters is proposed.In this approach,different positioning parameters are unified through graphical mapping,and the source is modeled as the key pixel in the image.By using the idea of key point detection,the deep neural network architecture of heterogeneous fusion positioning is designed to locate the source.Smulation results show that the proposed approach can flexibly fuse measurements of different types and quantities,achieving higher localization accuracy than benchmark algorithms.Secondly,considering that the measurement nodes are usually embedded with limited resources,this thesis studies the problem on how to balance the system overhead and positioning performance is studied,proposes a node selection algorithm based on Deep Q-Network.According to the dynamic nature of the source,the node selection problem is modeled as a Markov Decision Process(MDP),based on which a Q neural network is designed to fit the action value of the MDP.The neural network is trained based on the "try and observation" procedure,and the optimal selection is obtained based on the maximum value selection strategy.Simulation results show that when the number of nodes is small,the proposed algorithm can dynamically select nodes based on the movement of the source,and its performance is close to that of the exhaustive search algorithm.Finally,to address the problem that the node selection based on Deep Q-Network is not feasible when the number of nodes is large,a node selection algorithm based on improved DDPG is proposed.For the scenarios with a large number of nodes,a node selection MDP model for high-dimensional action spaces is designed.On this basis,an Actor network is designed to output node selection policies,and a Critic network is designed to fit policy values.The Actor and Critic networks are trained with the deterministic policy gradient algorithm.The continuous action output is mapped to a discrete action using the K-nearest neighbor algorithm and the maximum value selection strategy to obtain the optimal selection.Simulation results show that the proposed algorithm can dynamically select nodes in accordance with the movement of the source when the number of nodes is large,and its performance is superior to that of the nearest distance and random selection algorithms.