Spatial Feature Topology-Based Heterogeneous Knowledge Transfer Framework for Long-Term Fingerprint Positioning

Abstract

Transfer learning (TL) is effective for addressing distribution discrepancies in fingerprint positioning. However, existing TL frameworks cannot react well to the heterogeneous feature dimensions of fingerprints caused by the topology variations of base stations (BSs) in evolving long-term environments. To address this issue, we propose a spatial feature topology-based heterogeneous knowledge transfer framework tailored for long-term fingerprint positioning (SFTP). Firstly, we partition the heterogeneous features into common and domain-specific (including source and target-specific) features from static and dynamic environmental components perspectives, respectively. Notably, we observe that the features captured from BSs with significant spatial distances differ across all samples, while those from BSs with close spatial distances show higher similarity. Based on these observations, we approximate the cross-domain mapping for each domain-specific feature by integrating the mappings of similar common features, which are easy to achieve using deep neural networks (DNNs). Subsequently, the heterogeneous feature spaces are effectively transformed into homogeneous counterparts, and a deep adaptation network (DAN) is utilized to further predict the positions for testing samples. Hence, SFTP is capable of capturing evolutionary environmental information for long-term positioning. Finally, real-world experimental results demonstrate the superiority and robustness of the proposed framework.