Applying Transfer Learning on Various GNN Model Training in Indoor Positioning System Tasks

Abstract

Determining location and orientation has always been a fundamental challenge, driving advances from maps and compasses to modern global navigation satellite systems (GNSS). However, GNSS performs poorly indoors due to signal attenuation and lack of elevation accuracy, necessitating the development of indoor positioning systems (IPS). Various technologies such as Wi-Fi, Bluetooth Low Energy (BLE), and RFID have been deployed, typically relying on received signal strength (RSS) and fingerprinting to improve accuracy. While previous research focused on training a single model for an entire building, this study explores the creation of floor-specific models by applying transfer learning to various GNN models. This is done to address the substantial signal distortion between floors. Using the UTSIndoorLoc dataset, we evaluate Graph Attention Network (GAT), GraphSAGE, and Graph Convolutional Network (GraphConv) for predicting two-dimensional indoor positions based on RSSI fingerprints. We propose 2 transfer learning model training methods, Schema A and Schema B. Schema A trains the base model iteratively through each floor, and Schema B trains the base model on a unified dataset. Schema B with GraphConv achieved the best results with a mean positioning error of 6.2176 meters. Whilst Schema A achieved a best-case mean positioning error of 6.3900 meters. Both outperforming the standard unified model which has a mean positioning error of 8.0808 meters.