室内定位技术的广泛应用可极大地扩展基于位置服务的适用范围，能更好地满足个人用户、服务型机器人和智能物联网等日益增长的定位需求。考虑到基于位置指纹的室内定位问题可以被建模为监督学习任务，机器学习和深度学习已应用其中，但大部分这类方法并没有充分考虑到指纹数据潜在的非欧几里德数据特征。因此，本文通过构建图结构来表示指纹数据的此类特征，并采用图神经网络算法，研究基于Wi-Fi位置指纹和基于多源融合的室内定位技术。本文的主要工作如下：
（1）为了提升定位精度，考虑Wi-Fi接收信号强度（Received Signal Strength，RSS）之间的非欧几里德数据特征，改进了一种基于Wi-Fi的高阶图神经网络室内定位算法（Higher-order Graph Neural Network Indoor Localization Algorithm，HoGNNLoc）。将接入点（Access Point，AP）的空间几何信息纳入定位任务，以AP作为节点，对应RSS测量值作为节点特征，并根据AP位置间欧氏距离倒数设计邻接矩阵以构建图结构。基于此结构，应用高阶图神经网络聚合更新节点特征，并将更新后的节点特征拼接以形成完整的图特征表示，从而实现待定位目标位置的回归预测。在自建数据集和SoLoc公用数据集上进行实验，结果显示HoGNNLoc相较于对比定位算法具有更好的定位性能，可有效提升定位精度，并且具有较强的鲁棒性。
（2）针对Wi-Fi等单一信号源定位技术在动态环境中出现信号测量值丢失或出错的问题，考虑不同测量值数据之间的非欧几里德数据特征，改进了一种基于Wi-Fi、蓝牙和地磁多源信号融合的异构图神经网络室内定位算法（Heterogeneous Graph Neural Network Indoor Localization Algorithm，HeteroGNNLoc）。将不同类型的指纹数据转换为图数据，建立一个包含多种定位信息的异构图。为了表示不同节点之间存在的关系，利用训练数据集中的指纹数据统计信息为节点间添加边权重，丰富对定位信息的表达。并通过设定元路径区分不同类型的节点，构建异构图神经网络提取图数据特征，从而实现待定位目标位置的回归预测。在Miskolc IIS Hybrid IPS公用数据集上进行实验，结果表明HeteroGNNLoc相较于单一定位信号源定位算法显著提升了定位精度与稳定性，并且可以减轻构建指纹数据库的工作量。

The wide application of indoor localization techniques can greatly extend the applicability of location-based services and can better meet the growing localization needs of individual users, service robots, and smart internet of things. Considering that the indoor localization problem based on location fingerprints can be modeled as a supervised learning task, machine learning and deep learning have been applied, but most of such methods do not fully consider the potential non-Euclidean data features of fingerprint data. Therefore, this thesis investigates Wi-Fi location-based fingerprinting and multi-source fusion-based indoor localization techniques by constructing graph structures to represent such features of fingerprint data and employing graph neural network algorithms. The main work of this thesis is as follows:
(1) In order to improve the localization accuracy, a Wi-Fi based higher-order graph neural network indoor localization algorithm (HoGNNLoc) is improved by considering the non-Euclidean data features between Wi-Fi received signal strength (RSS). The spatial geometric information of access point (AP) is incorporated into the localization task, with AP as graph nodes and corresponding RSS measurements as node features, and the adjacency matrix is designed based on the inverse of the Euclidean distance between AP locations to construct the graph structure. Based on this structure, a higher-order graph neural network is applied to aggregate and update the node features, and the updated node features are spliced together to form a complete graph feature representation, thus realizing the regression prediction of the location of the target to be located. Experiments are conducted on the self-built dataset and the SoLoc public dataset, and the results show that HoGNNLoc has better localization performance compared with the comparison localization algorithm, which effectively improves localization accuracy, and the algorithm is more robust.
(2) Aiming at the problem of loss or error of signal measurements in the dynamic environment changes in single-source localization techniques, a heterogeneous graph neural network indoor localization algorithm (HeteroGNNLoc) based on the fusion of multi-source signals from Wi-Fi, bluetooth low energy and geomagnetism is improved to improve the localization accuracy and stability, while considering the non-Euclidean data characteristics between different measurements of the data. Different types of fingerprint data are converted to graph data to build a heterogeneous graph containing multiple localization information. In order to represent the relationship existing between different nodes, the statistical information of fingerprint data in the training dataset is utilized to add edge weights between nodes to enrich the expression of localization information. And by setting the meta-path to distinguish different types of nodes, the heterogeneous graph neural network is constructed to extract the features of the graph data, to realize the regression prediction of the location of the target to be localized. Experiments are conducted on the Miskolc IIS Hybrid IPS common dataset, and the results show that HeteroGNNLoc significantly improves the localization accuracy and stability compared with the single localization signal source localization algorithm, and reduces the workload of constructing the fingerprint dataset.