Adaptive WKNN Algorithm using Improved Distance and Extreme Learning Machine for Indoor Visible Light Positioning

In recent years, visible light positioning (VLP) using fingerprint technology has emerged as a competitive candidate for indoor positioning with high precision. However, the weighted K-nearest neighbors (WKNN) algorithm commonly used in fingerprint technology faces two major limitations. Firstly, the power difference used to measure the similarity between the target and reference nodes (RNs) does not accurately reflect the actual distance difference, leading to inconsistent weight allocation. Secondly, a fixed number of RNs (i.e., the K value) is not universally applicable to all targets, potentially resulting in reduced positioning accuracy for some. To overcome these limitations, we propose an adaptive WKNN algorithm using an improved distance metric and extreme learning machine (ELM). The improved distance metric is defined by applying logarithmic transformation to the power, making the power difference and distance difference approximately linear. Additionally, the improved distances between test targets and RNs, along with the optimal K values used to minimize positioning error, are pre-recorded. These data are then used to train the ELM network, enabling the online adaptive selection of the most suitable K value. Both software and hardware simulations show the effectiveness of the proposed algorithm in enhancing positioning accuracy while meeting real-time requirements.