基于机器学习的室内UWB三维定位

Indoor Ultra-Wideband 3D Positioning Based on Machine Learning

摘要: 随着无线通信网络的高速发展，对无线定位技术的要求越来越高。UWB是一种较为优秀的定位技术，但其容易受信号干扰，导致其在商场、酒店等室内场合的定位效果不够理想，故解决信号干扰下的UWB精确定位问题获得极大关注。本文使用的实验数据包含UWB定位的“无干扰数据”和“有干扰数据”。首先进行数据清洗，第一步根据箱线图找出数据中的异常值并剔除，第二步将数据进行K-means聚类，完成相同或相似数据的删除。然后对清洗后保留下的“无干扰数据”和“有干扰数据”使用遗传算法优化的BP神经网络算法(GA-BP)估计Tag的三维坐标，发现在GA-BP神经网络算法下，根据“无干扰数据”和“有干扰数据”得到的三维坐标的均方误差都非常接近于0，说明该模型预测精度较高。同时，根据GA-BP神经网络算法对10组测试数据进行精确定位，并建立XGBoost提升算法、支持向量机(SVM)和随机森林、KNN算法四种方法对数据是否在信号干扰下采集进行分类。通过比较四个模型的分类准确率和F-score，确定最终分类模型为经过参数优化的XGBoost算法。

Abstract: With the rapid development of wireless communication networks, the requirements for wireless positioning technology are getting higher and higher. UWB is an excellent positioning technology, but it is prone to signal interference, resulting in an unsatisfactory positioning effect in indoor places such as shopping malls and hotels. Therefore, solving the problem of UWB precise positioning under signal interference has received great attention. The experimental data used in this paper includes “no interference data” and “interference data” for UWB positioning. The first step is to clean the data. The first step is to find outliers in the data according to the boxplot and eliminate them. The second step is to perform K-means clustering on the data to delete the same or similar data. Then use the BP neural network algorithm (GA-BP) optimized by the genetic algorithm to estimate the three-dimensional coordinates of the Tag for the “undisturbed data” and “interference data” retained after cleaning. The MSE of the three-dimensional coordinates obtained from “interference data” and “interference data” is very close to 0, indicating that the model has high prediction accuracy. At the same time, according to the GA-BP neural network algorithm, 10 sets of test data are accurately positioned, and XGBoost is established. Boosting algorithm, SVM, random forest, and KNN algorithm are used to classify whether the data is collected under signal interference. By comparing the classification accuracy and F-score of the four models, it is determined that the final classification model is the parameter-optimized XGBoost algorithm.