Wi-Fi Fingerprinting Model Update For Indoor Localization Via Domain Adaptation

With the popularity of mobile devices and the large-scale deployment of the wireless access points(APs)in indoor environments,Wi-Fi RSS(Received Signal Strength)fingerprinting indoor localization has attracted much attention in the industry.In recent years,many outstanding works have come out in the filed of Wi-Fi RSS fingerprinting indoor localization,but there are still few that can be used in real-world scenarios yet.Because Wi-Fi fingerprinting indoor localization requires constructing a dense and accurate fingerprint database collected by professionals,which is timeconsuming and labor-intensive.However,Wi-Fi signals vary over time due to multipath fading and dynamic indoor environment.If the initial fingerprint database is used for long-run deployment of Wi-Fi fingerprinting localization,the localization accuracy will be significantly reduced.To retain high localization accuracy,the fingerprint database has to be update regularly,which is high-cost.How to reduce the cost of updating the fingerprint database and maintain the localization accuracy is still a serious challenge.In this paper,we proposed a novel unsupervised domain adaptation model TransLoc for Wi-Fi fingerprint update.The model is achieved only utilizing the initial fingerprint database and current unlabeled RSS fingerprints(without location information)to keep high accuracy at a low cost.TransLoc uses the theory of domain adaptation to learn the domain-invariant feature representation of the initial fingerprint database and unlabeled RSS fingerprints,in order to aligning the two RSS feature spaces.Therefore,domain adversarial training and cycle consistency loss are introduced to guarantee that the current and initial fingerprints at the same location are aligned in the domain-invariant feature space.At the same time,TransLoc learns how to utilize the domain-invariant features of labeled fingerprints to predict the location of the fingerprint.Then,thanks to the domain-invariant characteristics of the features,TransLoc can also predict the location on the current unlabeled fingerprints.On this basis,TransLoc introduces tri-training strategy to guarantee localization accuracy of the up-to-date unlabeled RSS fingerprints.In addition,in order to enable the model can converge to the best state faster,TransLoc also use the pre-training and joint training training strategy.To evaluate the performance of TransLoc,we carry out extensive experiments in two typical real-world indoor environments.The experimental environments are the corridors of the office building and the shopping mall.The initial fingerprint database is collected firstly and the current RSS fingerprints one month later in each environment.Experimental results show that,compared with other Wi-Fi RSS fingerprinting localization models and domain adaptation models,TransLoc has the least decrease in localization accuracy after one month,with a decrease of 15% and 18% in the office building and shopping mall respectively.In addition,we also set up a small area in the first scene,and measured Wi-Fi RSS for three months to verify the localization accuracy of TransLoc in the long run.The result shows that the decline of localization accuracy of TransLoc is the smallest accuracy compared with other localization methods after three months.In summary,TransLoc can effectively update the Wi-Fi fingerprinting localization model,maintain high localization accuracy at a low cost in the long run,and can be deployed and applied in large-scale scenarios.