A self-adaptive visible light indoor positioning system based on BLS network incorporating self-attention mechanism

Abstract

To address the challenges of degraded positioning accuracy and insufficient robustness caused by dynamic pose disturbances of the photodiode (PD) in visible light indoor positioning systems, this paper proposes a self-adaptive visible light indoor positioning system based on broad learning system (BLS) incorporating self-attention mechanism, referred to as SA-BLS. First, a visible light propagation model incorporating the 3D pose of the PD is established, and a multi-pose domain received signal strength (RSS) -pose multidimensional feature fingerprint database is constructed. Subsequently, a self-attention-BLS collaborative architecture is designed: the self-attention mechanism dynamically allocates weights to RSS and pose features, mitigating the RSS distribution shift caused by PD tilt, and achieves position estimation based on feature mapping matrices and pseudo-inverse analysis. Experimental results show that in a 4.5 m × 3.6 m × 3 m indoor environment, the traditional BLS model achieves average positioning errors of 5.49 cm and 19.97 cm under PD horizontal placement and tilt angles [0°, 30°], respectively. In contrast, the SA-BLS system maintains a stable average positioning error of 3.56 cm even when the PD is tilted within [0°, 30°]. These results validate the significant advantages of the proposed system in improving positioning accuracy and adaptability, demonstrating strong potential for practical applications.