面向煤矿井下应用场景的指纹定位算法研究

Underground positioning system which is directly related to the safety of mine workers, is an important part of the mine refuge system. However, the popular fingerprint-based localization algorithm has some defects, such as the fingerprint can not reflect the underground environment characteristics accurately and completely, the matching algorithm is inaccurate and time-consuming, etc.. On the basis of the our previous work on the subject, we apply the deep learning framework, which is recently most attractive in the information processing area, to the special environment of the underground tunnel in this project. Our research includes how to find the fingerprint database based on ray tracing method and channel state information, and how to match the fingerprint to the reference location based on deep learning. We will focus on the following two key scientific issues: (1) How to reflect the complex multipath environment with the fingerprint data? (2) What are the main factors which affect the accuracy and effectiveness of fingerprint matching? One of the objective of the project is to find a novel multpath based fingerprint database from both theoretical channel model and real data，which could appropriately reflect the characteristics of the channel cross section and the device and equipment in the channel. The other one is to propose a new matching algorithm based on deep learning, we will set the input, output of the deep network, as well as the number of hidden layers. Moreover we will find a reasonable training and fine-tuning scheme as well as data fusion based location estimation. The research result of this project will provide a new idea for the design of underground positioning system, and will also provide theoretical and talent reserve for its industrialization.

井下人员定位系统直接关乎井下从业人员的生命安全，是井下安全避险的重要一环，但目前流行的基于指纹的定位算法存在着指纹信息不能准确反映井下环境多变特性，匹配算法不精确、耗时长等问题。本项目在前期工作的基础上将信息处理学科中最新发展的深度学习框架应用于井下巷道特殊环境中，研究基于射线追踪法和信道状态信息的指纹库建立方法以及基于深度学习的指纹匹配算法。项目将围绕井下复杂多径环境与指纹库数据之间的映射关系和深度学习框架中的哪些因素会影响指纹匹配的精确性和速效性两个关键科学问题展开。旨在从理论模型和实际数据两方面建立起能反映巷道横截面特征、巷道内障碍等基本影响因素的指纹映射新模式；提出利用深度学习框架的指纹匹配新算法，包括明确深度学习模型的输入、输出和隐层数，确立训练和调优机制和在线快速融合准则。本项目的研究内容将对井下人员定位系统的设计提供新思路，同时为它的产业化工作提供理论和人才储备。

目前井上常用的精密定位算法无法直接移植到井下，迫切需要研究适合于井下环境特点的高精度快速定位算法。本项目主要针对目前井下指纹定位算法存在的几个问题展开，一个是更为准确的指纹库建立方法，另一个是快速精确的指纹匹配算法。所取得的主要成果如下：.以射线追踪法为基础，建立了矿井巷道电磁波传播模型，并且仿真得出随着收发端距离的增大，接收功率、路径损耗的变化趋势。对得到的传播模型进行仿真，根据仿真结果分析巷道横截面、频率、天线位置、极化方式、巷道壁粗糙度以及巷道岩层电参数等因素对矿井巷道电磁波的接收功率和路径损耗的影响。根据建立的电磁波传播模型获得各个参考点处每一条射线的功率强度作为指纹样本，并对其进行处理，建立数据库，为之后精确的的定位提供数据支持。.对移动终端采集的真实CSI幅度值用密度最大值聚类算法进行预处理，剔除了异常值，构建了位置指纹数据库。对比了网格训练单元大小，天线数目，学习算法等几个参数对被动定位的影响。提出了一种使用严格设计的人工神经网络的指纹数据库更新方案，旨在对室内结构和布局变化引起的定位误差进行建模，降低指纹库更新过程中的人工成本。.使用机器学习中稀疏表达理论（Sparse representation theory）相关方法，将代表井下不同地点的指纹信息构建有效编码字典，分别对不同采样点进行多标签编码学习，深入挖掘了CSI信息的内部特质，方便后续定位算法处理过程。.上述研究成果从井下无线信道建模、WiFi信道状态信息分析、基于深度学习框架的匹配算法等角度出发，研究一套完整的井下指纹定位算法，为井下人员定位方案提供了理论指导。