Accurate localization is one of the basic requirements for smart cities and smart factories. In wireless cellular network localization, the straight-line propagation of electromagnetic waves between base stations and users is called line-of-sight (LOS) wireless propagation. In some cases, electromagnetic wave signals cannot propagate in a straight line due to obstruction by buildings or trees, and these scenarios are usually called non-LOS (NLOS) wireless propagation. Traditional localization algorithms such as TDOA, AOA, etc., are based on LOS channels, which are no longer applicable in environments where NLOS propagation is dominant, and in most scenarios, the number of base stations with LOS channels containing users is often small, resulting in traditional localization algorithms being unable to satisfy the accuracy demand of high-precision localization. In addition, some nonideal factors may be included in the actual system, all of which can lead to localization accuracy degradation. Therefore, the approach developed in this paper uses knowledge graph and graph neural network (GNN) technology to model communication data as knowledge graphs, and it adopts the knowledge graph inference technique based on a heterogeneous graph attention mechanism to infer unknown data representations in complex scenarios based on the known data and the relationships between the data to achieve high-precision localization in scenarios with LOS/NLOS channel coexistence. We experimentally demonstrate a spatial 2D localization accuracy level of approximately 10 meters on multiple datasets and find that our proposed algorithm has higher accuracy and stronger robustness than the state-of-the-art algorithms.

精准定位是智慧城市、智慧工厂的基本要求之一。在无线蜂窝网络定位中，电磁波在基站和用户之间的直线传播称为视距（LOS）无线传播。在某些情况下，电磁波信号由于建筑物或树木的遮挡而无法沿直线传播，这些场景通常称为非视距（NLOS）无线传播。传统的定位算法如TDOA、AOA等基于LOS信道，在非视距传播占主导地位的环境中不再适用，而且在大多数场景下，包含用户的LOS信道的基站数量往往较少，导致传统的定位算法无法满足高精度定位的精度需求。此外，实际系统中可能存在一些非理想因素，这些因素都会导致定位精度下降。因此，本文提出的方法利用知识图谱和图神经网络（GNN）技术将通信数据建模为知识图谱，并采用基于异构图注意力机制的知识图谱推理技术来推理复杂场景中的未知数据表示基于已知数据以及数据之间的关系，实现LOS/NLOS信道共存场景下的高精度定位。我们在多个数据集上通过实验证明了大约 10 米的空间二维定位精度水平，并发现我们提出的算法比最先进的算法具有更高的精度和更强的鲁棒性。