Thickness sensing and anomaly detection for shotcrete support in roadways based on machine learning

Abstract: Aiming at the problem that the shotcrete thickness detection in roadway engineering relies on discrete manual spot checks and it is difficult to identify continuous low-thickness regions, this paper constructs the Full-field Thickness Perception and Anomaly Detection Method (FTP-AD). Taking the two-phase LiDAR point clouds before and after shotcreting as input, this method establishes the surface correspondence through global registration and ROI local registration, and uses floor correction and principal normal translation preservation to reduce the offset of registration to the change of shotcrete thickness. Subsequently, while retaining the candidate points at high-curvature boundaries, sparse outliers are eliminated, a finite-length normal cylindrical neighborhood is used to estimate the local thickness, and the thickness field is formed through median smoothing, then the connected low-thickness regions are extracted combined with the designed thickness limit. Verified by 20 measuring points on a single roadway interface, the results show that the mean absolute error of FTP-AD is 4.52 mm and the root mean square error is 5.69 mm, both lower than the comparison configurations using traditional ICP, point-to-plane ICP and M3C2. The ablation results further indicate that both floor correction and principal normal translation preservation help to reduce the thickness measurement error. In this case, the proposed method realizes the spatial expression of the shotcrete thickness distribution and the concentrated low-thickness positions, and provides quantitative information for the quality evaluation of roadway shotcrete support.