A Deep Learning-Based Method for LOS/NLOS Detection in Harsh Underground Mining

Accurate identification of line-of-sight (LOS) and non-line-of-sight (NLOS) conditions is a key factor in improving the performance of wireless localization systems in underground mining environments, where signal degradation caused by obstructions and multipath effects is significant. This paper presents a novel approach that combines a carefully selected set of channel impulse response (CIR) parameters with a lightweight convolutional neural network (CNN) for reliable LOS/NLOS classification of Wi-Fi signals. The proposed method converts CIR features into compact image representations, enabling efficient processing while preserving essential discriminative information. The selected features include statistical, spectral, and amplitude-based parameters designed to capture differences in signal structure between LOS and NLOS conditions. The model is trained and validated using real-world measurements collected in an operational underground mine. Experimental results are obtained using three MIMO antenna configurations: planar, conformal, and Butler. The proposed approach achieves excellent classification performance, with accuracies of 100% for planar and Butler antennas, and a true negative accuracy of 97.2% for the conformal antenna. These results demonstrate the effectiveness of the proposed CIR feature combination and emphasize the importance of appropriate feature selection for improving CNN-based LOS/NLOS classification in complex underground environments.