Graph Neural Network based Localization for Layout-Agnostic Sensor Networks

This paper addresses the collaborative localization problem in sensor networks based on distance measurements, focusing on layout-agnostic sensor networks where the number and positions of sensor nodes to be localized can vary dynamically. To tackle this challenge, we propose a novel Graph Neural Network (GNN)-based localization algorithm, named Message Passing Neural Network Localization (MPNN-Loc), which integrates a message-passing mechanism with a correlation-based refinement strategy. This approach effectively extracts intrinsic information from the network structure and mitigates the impact of noisy measurements. Once trained, the MPNN-Loc algorithm can be directly applied to localize sensor nodes in layout-agnostic networks with varying numbers and positions, as long as a small subset of nodes, referred to as anchor nodes, have fixed and known positions. Extensive evaluations on layout-agnostic sensor networks with diverse configurations demonstrate that our proposed algorithm exhibits excellent generalization ability and strong robustness. These results highlight the potential of MPNN-Loc for real-world applications in dynamic and uncertain environments.