VNP: A Weak-Supervised and Retraining-Free Learning Approach for Online Positioning

The use of user equipment (UE) for positioning offers many benefits in indoor and urban areas, such as low latency and easy integration. However, acquiring accurate labels and achieving network generalization can be challenging. To address these issues, we propose a novel network architecture called variational neural processes (VNP), which is free from either precise labels or network retraining. VNP combines the advantages of variational inference and stochastic process regression theory. By generative model, environmental sensing is realized in weak-supervised manners. By neural processes, NLOS ranging bias can be always derived from environment-related information. Simulations demonstrate that our approach achieves high-precision positioning without requiring precise labels or network retraining, and it outperforms conventional techniques in terms of both positioning accuracy and generalization capability.