The accuracy of positioning in ultra-wide band (UWB) systems can be improved by identifying the conditions of line-of-sight (LOS) and non-line-of-sight (NLOS) propagation and taking appropriate measures. At present, machine learning methods have already been extensively applied to NLOS identification. Unfortunately, those traditional methods are incapable of performing well in the cases of small samples and class-imbalance samples. In this letter, a novel method is proposed to identify LOS and NLOS components by utilizing Capsule Networks (CapsNet). As indicated by simulation results, the CapsNet-based method allows LOS and NLOS measurements to be classified with 94.63% accuracy, 95.58% precision, 94.74% recall rate, and a 0.9490 F2-score. In addition, CapsNet is compared against decision tree (DT), least squares vector machine (LS-SVM) and K nearest neighbor (KNN) based on experimental data. The comparison results show that the proposed CapsNet outperforms benchmark methods in the cases of both small samples and class-imbalance samples.

中文翻译：

---

使用胶囊网络进行 UWB 定位的非视距识别


通过识别视距（LOS）和非视距（NLOS）传播条件并采取适当措施，可以提高超宽带（UWB）系统中的定位精度。目前，机器学习方法已经广泛应用于非视距识别。不幸的是，这些传统方法在小样本和类不平衡样本的情况下无法表现良好。在这封信中，提出了一种利用胶囊网络（CapsNet）来识别 LOS 和 NLOS 组件的新方法。仿真结果表明，基于 CapsNet 的方法可以对 LOS 和 NLOS 测量进行分类，准确率达 94.63%，精确度达 95.58%，召回率达 94.74%，F2 分数达 0.9490。此外，基于实验数据，CapsNet 与决策树（DT）、最小二乘向量机（LS-SVM）和 K 最近邻（KNN）进行了比较。比较结果表明，所提出的 CapsNet 在小样本和类不平衡样本的情况下都优于基准方法。