Precise multi-dimensional features positioning of Xianglushan tunnel drilling based on deep-machine vision
[Objective]Xianglushan tunnel is a key project of the Yunnan Dianzhong Water Diversion Engineering,and rock bursts are one of its main hazards.The precise positioning of microseismic sensor coordinates is the basis for microseismic monitoring and early warning systems for rock bursts,which is controlled by drilling coordinates and angles.However,the drilling location needs to be measured manually,which has low accuracy and poor timeliness,seriously hindering the development of automatic microseismic monitoring systems for rock bursts.Meanwhile,the precise positioning of multi-dimensional features is the key to realizing the full automation of drilling operations,including the coordinates,diameter,plane direction,and busbar direction of drilling.This study presents a method for the precise positioning of multi-dimensional features of tunnel drilling based on deep-machine vision called YOLO-AT.[Methods]The proposed method includes two modules:drilling contour detection and drilling multi-dimensional feature positioning.The drilling contour detection module adds rotation angle prediction to the YOLO V8 model,thereby constructing a YOLO V8 OBB model that can accurately fit the drilling contour.The drilling multi-dimensional feature positioning module establishes an anchor tracking algorithm,selects anchor points on contour ellipses,and achieves anchor tracking from multiple perspectives based on epipolar line constraints and projective invariance of anchor order.The spatial coordinates of each anchor point and multi-dimensional features were solved using the parallax method and analytical geometry theory,respectively.[Results]To verify the robustness of the proposed method to contour quality,a synthetic drilling contour database with controllable quality was established for Xianglushan tunnel,and the positioning results were compared with those of the ALSR,FED,AAMED,YOLO V8,and DLT methods.Results showed that:(1)concerning contour detection accuracy,the proposed method exhibited the best performance among other methods,with the average intersection over union(IoU)ratio,average F1 score,precision,and recall at an IoU threshold corresponding to 0.9 of 0.957,0.948,0.977,and 0.977,respectively.(2)Concerning contour detection stability,the proposed method was less affected by contour quality and always maintained high accuracy.By contrast,YOLO V8 had less fluctuation and low indicators,whereas other methods had poor stability and high accuracy under good contour quality.(3)Concerning the accuracy of multi-dimensional feature positioning,the median errors of the proposed method for the coordinate,diameter,plane direction,and busbar direction of drilling were 0.835 mm,0.795 mm,0.567°,and 1.751°,respectively,which were the best among all methods.(4)Concerning the stability of multi-dimensional feature positioning,the proposed method and YOLO V8-AT had better stability.Conversely,the stability of other methods rapidly decreased as the quality of the drilling contour deteriorated.[Conclusions]The proposed Y OLO-AT method can achieve accurate and stable positioning of multi-dimensional features in Xianglushan tunnel drilling,with a performance superior to that of existing methods,thereby having the potential to be extended to various underground engineering drilling positions.Next,the research intends to focus on combining the proposed method and laser ranging technology to measure hole depth and eliminate ambiguity in drilling direction,in addition to analyzing the roles of RANSAC and reprojection methods in improving algorithm accuracy and stability.
deep learningmachine visionmicroseismic monitoringdrilling positioningDianzhong water diversion engineering
NETL
NSTL
万方数据
