摘  要

室内定位技术可以为基于位置的服务（LBS）带来更大范围的扩展，同时也为个人用户、服务型机器人和智能物联网装置提供更多的定位需求。根据室内定位原理，将其划分为几何测量和位置指纹两大类。由于传统的几何测量在非视距（NLOS）情况下定位性能受限，因此，位置指纹法受到了更多研究者的关注。考虑到位置指纹法可以被建模为监督学习问题，机器学习与深度学习在处理这类问题时表现优异，因此，本文采用机器学习和深度学习算法，研究基于Wi-Fi位置指纹的室内定位技术。本文的主要工作和创新点如下：

（1）为了解决在动态变化的环境中定位精度不高的问题，本文提出一种基于XGBoost结合弹性网实现误差补偿的算法（XGB-ENEC）。首先使用XGBoost算法对待定位点进行初步定位，再使用弹性网算法对当前预测的结果与真实坐标的差值进行预测，以此来修订初步定位的结果，实现误差补偿。该算法通过对接入点（AP）加噪，模拟室内环境的动态变化，采用现场测量的数据集进行仿真。实验结果表明，在80%分位处，XGB-ENEC算法的定位精度能够达到0.75m，明显优于其他定位算法。因此，XGB-ENEC算法可用于在动态变化的室内环境实现高精度定位。

（2）为了提高定位精度，将深度学习中的自编码器（AE）应用于室内定位。在处理位置预测问题时，将其建模为回归问题，通过改进损失函数并设计基于AE的位置预测模型。在仿真数据集上对模型进行训练和测试，实验结果表明，基于AE的位置预测模型能实现较高的定位精度，其预测轨迹与实际轨迹拟合度较高。

（3）针对UJIndoorLoc数据集，为了对高维数据降维，提高建筑物和楼层的分类准确率，将其建模为分类问题。通过改进损失函数，添加AE输出端与真实标签的交叉熵损失函数，并设计基于AE的建筑物楼层分类模型。实验结果表明，在测试集上，分类准确率可以达到97.55%，优于其他机器学习算法。以上实验结果表明，AE算法能够有效提取数据的主要特征，具有强大的非线性映射功能，用于位置预测时可以有效提高定位精度，用于建筑物楼层分类时可以有效提高其分类准确率。

关  键  词：室内定位；Wi-Fi；位置指纹；误差补偿；自编码器

研究类型：理论研究

ABSTRACT

Among indoor positioning technologies, location based services (LBS) can expand the field of application, and it also helps individual users, service robots and generous intelligent internet of things (IoT) devices meet the positioning needs. Indoor positioning methods can be divided into two categories: geometric measurement method and location fingerprint method. Due to the low positioning accuracy of traditional geometric measurement methods in non line of sight (NLOS) cases, the researchers focus on indoor positioning based on location fingerprint method. Considering that location fingerprinting can be modeled as a supervised learning problem, machine learning and deep learning perform well in dealing with such problems. Therefore, this paper adopts machine learning and deep learning algorithms to study the indoor positioning technology based on Wi-Fi location fingerprint. The main work and innovations of this paper are as follows:

(1) For the limited localization accuracy in the case of dynamic changes, the error correction of the localization results is performed by using XGBoost in combination with elastic networks (XGB-ENEC). First, XGBoost algorithm is used for initial location of anchor points, and then elastic network algorithm is used to predict the difference between the current predicted results and real coordinates, so as to revise the initial location results and realize error compensation. The algorithm adds noise to access point(AP) to simulate a dynamic change in the indoor environment, and the experiments are carried out on the measured dataset. Experimental results show that using the XGB-ENEC algorithm, sample positioning accuracy of 80% data can reach 0.75m, which is obviously better than other positioning algorithms. It is proved that XGB-ENEC algorithm can achieve high precision positioning in dynamic indoor environment.

(2) In order to improve the positioning accuracy, the autoencoder (AE) in deep learning is applied to indoor positioning. When dealing with the position prediction problem, it is modeled as the regression problem, and the position prediction model based on AE is designed by improving the loss function. The model is trained and tested on the simulation dataset, and the experimental results show that the position prediction model based on AE can achieve high positioning accuracy, and the predicted trajectory has a high degree of fitting with the actual trajectory.

(3) Aiming at UJIndoorLoc dataset, in order to reduce the dimension of high-dimensional data and improve the classification accuracy of buildings and floors, this problem is modeled as the classification problem. By improving the loss function, the cross entropy loss function of AE output end and real label is added, and the building floor classification model based on AE is designed. Experimental results show that the classification accuracy can reach 97.55% on the test set, which is better than other machine learning algorithms. The above experimental results show that AE algorithm can effectively extract the main features of data, has a powerful nonlinear mapping function, and can effectively improve the positioning accuracy when used for indoor location prediction, and can effectively improve the classification accuracy when used for building floor classification.

Key words：Indoor positioning; Wi-Fi; Location fingerprint; Error compensation; Autoencoder

Thesis          ：Theoretical research