A Study Of CSI-based Passive Localization Method Adapted To User And Environmental Changes

In recent years,with the miniaturization of computing devices and the intelligence of home products,indoor location services have become increasingly important as the basis for many smart services.WIFI-based channel state information fingerprint localization services have attracted much attention from researchers by virtue of the wide deployment and low cost of related routing devices.However,existing CSI fingerprinting methods are usually based on static environment assumptions,ignoring the problem of accuracy stability under user or environment change scenarios.Specifically,when the user identity or indoor environment changes,the fingerprint data distribution will also change due to the change of signal reflection and diffraction paths,making it difficult for traditional models to complete high-precision localization.Moreover,the perturbation of fingerprint distribution caused by the change of indoor environment is more drastic than the change of user identity and has higher diversity.Therefore,it is very difficult to design a stable feature analysis strategy for the above scenarios using CSI channel features to achieve passive localization under user identity or indoor environment change scenarios.The main research works and contributions of this thesis are as follows:(1)To address the problem of degraded recognition performance of the localization model caused by inconsistent CSI fingerprint data due to user identity changes in WIFI fingerprint localization systems,this thesis proposes a localization method based on generating samples to expand the fingerprint library.Firstly,the fingerprint library is expanded by generating artificial samples by learning the feature distribution of real samples.In this thesis,we use the βVAE-C model to decouple the features of CSI data,learn the complex feature distribution and fit it to a normal distribution parameterized by the available mean vector to generate more realistic artificial samples.On this basis,the amplitude and phase difference data in the CSI are extracted to form RGB images as fingerprints,and the resolution of the fingerprints is enhanced by the rich channel information.Experiments are conducted in two different indoor environments,and the experimental results show that the proposed method achieves 98.7% and 98.2% localization accuracy for tagged users,and 96.51% and 86.53%localization accuracy for new users not in the fingerprint database,respectively.This verifies that the passive WIFI localization method in this thesis has better performance.(2)To address the problem of degradation in the recognition performance of the localization model caused by environmental changes,this thesis proposes a CSI fingerprint localization model based on sub-domain distribution alignment from the perspective of domain adaptation.The core of the problem of user and environment changes is the change of domain information in the data distribution.To address this problem,this thesis first calculates the reconstruction loss of samples before and after the change of domain information and extracts the domain-independent common features.Based on this,we use the decision results of unlabeled data as soft labels for semi-supervised learning,and calculate the maximum mean difference of the data before and after the domain information change on the fingerprint of each locus marker to align the feature distribution and improve the generalization ability of the model.Finally,this thesis conducts experiments for both environmental changes as well as user changes,respectively,four different indoor environments and four different user scenarios,and the results show that the overall recognition accuracy is improved by about 5% and 12% compared with the unsupervised global domain adaptation classification algorithm,and about 8% and 16% compared with the traditional localization scheme,which has better localization performance and effectively reduces the resource waste of unlabeled samples.