Wi-Fi Perception by Using Multiple Kernel-Based Learning for Robust Positioning

作者：Li X, Hongkun Du, Yunqi Jia, Lu Xu, Shenjie Tang · 发表于：IEEE Transactions on Vehicular Technology · 年份：2026 · DOI：10.1109/tvt.2026.3693443 · 研究领域：Indoor and Outdoor Localization Technologies、Direction-of-Arrival Estimation Techniques、Wireless Networks and Protocols

The expansion of the Internet of Things has driven significant growth in wireless applications. Advancements in Wi-Fi and its evolution have raised the importance of Wi-Fi based localization. However, contemporary machine learning approaches often lack the ability to dynamically capture reliable features in real-world situations where signal fluctuations exist, resulting in reduced robustness and accuracy. To address this issue, we introduce an information-theoretic method for robust and precise Wi-Fi based localization under uncertainty. Concretely, we adopt a Laplace-Student's$t$mixture correntropy (LTMC) error model to boost accuracy and robustness, coupled with an enhanced entropy-K-means Nyström (EEK-Nyström) sparse model to limit neural network expansion. The process consists of two stages: offline training and online localization. In the offline stage, we construct a mixed-kernel correntropy model employing both Laplace and Student's$t$kernels as the cost function and derive its sparse variant via an entropy-enhanced Nyström method. In the online stage, we use the mixed correntropy model to evaluate fingerprint similarity and apply a mixed-correntropy-driven adaptive filtering algorithm to enhance localization precision. Experimental results demonstrate that the proposed algorithm delivers superior accuracy and robustness, achieving an average error of 1.15 m under real-world disturbances following impulsive and Gaussian distributions. By utilizing existing Wi-Fi syste...