A fractional-derivative kernel learning method for indoor position prediction

Precise indoor positioning is increasingly essential for Internet of Things (IoT) deployments. Within the IoT framework, many resource-limited devices require accurate positioning capabilities. Therefore, as a lightweight online learning algorithm, kernel adaptive filter (KAF) has been adopted for indoor positioning. However, the environments and mixed noise often lead to the inaccuracy of indoor positioning. In order to tackle this problem, this paper proposes a novel fractional-derivative kernel adaptive filtering method, which has two significant features: The fractional-derivative assisted weight-update strategy can accurately reflect data distributions and system characteristics, thereby improving the accuracy of prediction. Secondly, the Student's t-based kernel function can effectively combat non-Gaussian noise in indoor positioning applications, thus the prediction robustness can be well guaranteed. In three representative indoor positioning scenarios, the proposed algorithm is compared with state-of-the-art KAFs and widely used positioning methods (Trilateration, KNN, Naive Bayesian, and transfer learning). Experimental results show that the proposed method can improve positioning accuracy by at least 5.98% with the well-robustness property.