Hybrid CNN and XGBoost model for non-line-of-sight classification in Ultra-Wideband systems
Аннотация
Ultra-Wideband (UWB) positioning systems are used in indoor localization and intelligent sensing applications because of their high ranging precision in complex environments. However, Non-Line-of-Sight (NLOS) propagation remains a major challenge affecting localization accuracy. This paper first reproduces the deep learning framework proposed by Hwang et al. for UWB LOS/NLOS classification using Channel Impulse Response (CIR) signals. The reproduction follows the same dataset, preprocessing pipeline, and evaluation metrics reported in the original study. The reproduced Transformer-CNN achieves performance close to the reported results, providing a consistent reference for subsequent evaluation. After reproducing the baseline, this paper investigates a hybrid learning framework that combines raw CIR waveform learning with engineered signal-domain feature learning. The framework integrates a CNN trained on raw CIR signals with an XGBoost classifier trained on engineered UWB features. A Logistic Regression meta-classifier combines the prediction probabilities of both branches through stacking. Experimental results show that the hybrid framework achieves an accuracy of 90.12% and an AUC of 0.9688, compared with 86.07% accuracy and 0.9333 AUC for the reproduced Transformer-CNN. The standalone XGBoost model achieves 89.98% accuracy and .9671 AUC, indicating that the additional improvement obtained through hybrid fusion is modest. Additional evaluation includes multi-seed stability, component-level and engineered-feature ablation analyses, Gaussian-noise robustness testing, prediction-correlation and SHAP-based interpretability analyses, confidence interval estimation, and McNemar statistical significance testing.