UWB and Inertial-Measurement-Driven Res2Net-Triplet Model for Mine-Roadway Personnel Positioning Correction
Authors
Kovács István, Nagy Péter, Farkas TamásFiles
Abstract
Accurate positioning of underground personnel is still challenging due to non-line-of-sight propagation of UWB ranging, metallic equipment, irregular tunnel structures, and inertial drift caused by walking. This paper proposes a UWB-and-inertial-measurement-driven Res2Net-Triplet model for mine-roadway positioning correction to improve the average accuracy and tail-error control without modifying the deployed ranging infrastructure. Convert synchronous UWB ranges, channel-quality descriptors, accelerometer windows, gyroscope windows, anchor-visibility masks and coarse coordinates into a multi-scale temporal representation. Res2Net blocks capture short disturbance spikes and wider motion context, and triplet metric learning separates correction patterns caused by open-view walking, partial shielding, turning, and severe NLOS disturbance. A lightweight regression head estimates coordinate offsets and does not replace the physical localisation solution. In a simulated roadway dataset with 18 routes, 42 anchor configurations and 12,600 labeled trajectory windows, the proposed model reduced the three-dimensional mean absolute positioning error from 0.94 m to 0.31 m and decreased the 95th-percentile error from 2.18 m to 0.76 m. Compared with EKF, LSTM and a plain residual CNN baseline, the method improves F1 detection of high-risk error windows to 0.91, reduces the mean error of turning-NLOS to 0.38 m, and keeps the median correction latency at 18.7 ms per window. The results show that multi-scale residual features and distance-constrained embeddings can address the problems of altered anchor geometry, body shielding and inertial drift along the route practically.