Online Transfer Learning Theories and Methods for Fingerprint-based Indoor Positioning

Abstract:

Accurate location services are essential in the Internet of Things (IoT) era because

location-based services (LBS) rely on the user's location data to provide contextual

functionality. The complexity of indoor environments limits the use of Global Positioning System (GPS) signals in indoor positioning systems (IPS), despite the fact that they are widely accepted as the most reliable and promising method of determining location in outdoor environments. In indoor positioning systems, transfer learning (TL) techniques are often used to predict the location of mobile devices, presuming that all training instances of the target domain are provided beforehand. In dealing with the issue of changing signal distribution, particularly in dynamic indoor environments, this assumption has been criticized for its shortcomings. This could justify the possibility of training instances arriving online, feature spaces of the source domain and target domain may differ, the expense of gathering a large enough pool of training instances is costly, or the possibility that negative knowledge might be transferred in the event that

the source domains are redundant.

Thus, a novel online heterogeneous transfer learning (OHetTLAL) algorithm of IPS-based fingerprinting is proposed to improve the localization performance of the target domain by fusing the knowledge of the source domain and target domain. The source domain is refined according to the target domain to avoid negative knowledge transfer. The co-occurrence measure of the feature spaces was used to derive the homogeneous new feature spaces, and the features with higher weight values in relation to the instances of the target domain were selected for training the classifier because they could positively affect the location prediction of the target. In addition, a novel heterogeneous transfer learning (HetTL) method based on hybrid feature selection has been applied to reduce training calibration effort and the noise generated by fingerprint duplications obtained from multiple Wi-Fi access points. Although location services have come a long way over the years, there is still a need for an alternate technology for indoor positioning since indoor signals are more vulnerable to NLOS propagation effects, multipath effects, and a dynamic environment than outdoor signals. Thus, a positive knowledge transfer-based heterogeneous data fusion method is proposed for representing the different scenarios of temporal variations of CSI-based fingerprint measurements generated in a complex indoor environment targeting indoor parking lots, while also reducing training and calibration overhead. Furthermore, a CRLB analysis was also used to estimate the lower bound of the parking lot location error variance under various temporal variation scenarios and establish some geometric relationships to describe the location estimator.

Moreover, this dissertation analyzes various measurement techniques and technological solutions for dealing with complex phenomena in indoor environments. It mainly consists of four parts: i) Online Heterogeneous Transfer Learning (OHetTLAL) algorithm for IPS-based fingerprinting to improve localization performance in the target domain by fusing source and target domain knowledge; ii) comparative analysis on theories and methods of different IPS wireless technologies based on different approaches; iii) challenges in wireless indoor localization scenarios; and iv) prospective measures have been presented as future research directions. The dissertation, in particular, examines and assesses a number of existing positioning systems based on multidimensional matrices of various architectural and design problems and suggests effective IPS considering the various evaluation metrics described in the literature. In contrast to GPS, which is the industry standard for outdoor locations, it also discusses the challenges that limit the deployment of IPS and the feasibility of them becoming critical infrastructure. Furthermore, it outlines some prospective measures that can be taken to address the challenges posed by indoor positioning. These methods can be used to enhance the overall estimation and localization performance of IPS, including transfer learning methods, feature engineering techniques, data fusion methods, multi-sensory techniques, hybrid techniques, and ensemble learning.

Keywords: Indoor Positioning, Online Transfer Learning, Machine Learning Algorithms, Heterogeneous Features, Optimization

Final_PhD_Dissertation_Tesfay_Gidey_Hailu_2023.pdf

Download PDF • 4.66MB