Research On Transit Time Model Of UWB On Gaussian Process And Transfer Learning

The accuracy of ultra-wideband indoor positioning mainly depends on NLOS error and multipath interference caused by complex environment.In the NLOS environment with occlusion,the ranging accuracy will deviate and the positioning effect will be greatly reduced,so reducing NLOS error is one of the key points of research.Based on the arrival time and the time difference of arrival are commonly used algorithms to calculate the location,so the accuracy of the UWB signal transmission time between the anchor node and the node is very important.The time loss in different environments is different.How to use less manpower to calculate the ideal signal transmission time is also important.To solve these problems,Thesis introduces a nonparametric NLOS target location strategy.First of all,the ideal transmission time of UWB signal transmission under non-line-of-sight conditions and the transmission time in the actual environment are collected for training the GPR model.The DE algorithm searches the super parameters of the GPR model,and establishes the DE-GPR model.Then,the DE-GPR return is used to generate the estimation of the transmission time of the whole area to be located under ideal conditions.If the location environment changes,the original model cannot be fully applicable in the new environment.Thesis uses the improved TrAdaBoost algorithm based on migration learning to filter the original modeling data,which can reduce the impact of data with large errors and retain the data still applicable in the new environment.This greatly reduces the time and cost of modeling in the new environment,and enhances the generalization ability of the model,The specific contributions of Thesis are as follows.1.A time loss propagation model DE-GPR based on Gaussian Process is proposed.The algorithm uses DE algorithm to select the optimal hyperparameter for the GPR model,takes the actual transmission time as the input,and the theoretical transmission time as the output.The trained model can directly predict the time information we need for positioning.2.A new TrAdaBoost-DEGPR algorithm based on transfer learning is proposed.The classification model in the traditional TrAdaBoost algorithm is replaced by a regression model,and the DE-GPR model proposed in Thesis is used as the basic regression.Combining the DE-GPR model with the migration learning technology,continue to use the expired data,retain the data that is still applicable to the new environment,and reduce the time and labor costs required for measuring data in the new environment.3.Through a large number of simulation experiments,the effectiveness and robustness of the proposed scheme are verified,and competitive positioning accuracy can be achieved.