| 肖尧,钟登华,余佳,胡奕可,徐国鑫,陈秋同.基于改进HT-LCNN线段检测模型的隧洞施工活动时间信息智能提取方法[J].水利学报,2024,55(1):24-34,47 |
| 基于改进HT-LCNN线段检测模型的隧洞施工活动时间信息智能提取方法 |
| An intelligent extraction method of tunnel construction activity duration information based on improved HT-LCNN |
| 投稿时间：2023-05-18 |
| DOI：10.13243/j.cnki.slxb.20230288 |
| 中文关键词: 隧洞施工活动时间 信息智能提取 线段检测 深度霍夫线条先验网络 注意力机制 |
| 英文关键词: duration of tunnel construction activities information intelligent extraction line segment detection deep network with Hough transform attention mechanism |
| 基金项目:国家自然科学基金项目(52279137，52009090) |
|
| 摘要点击次数: 1447 |
| 全文下载次数: 682 |
| 中文摘要: |
| 施工活动时间信息的有效提取是隧洞施工进度分析与决策的重要前提，目前常采用的一种隧洞施工活动时间记录方式是绘制横道图线段。针对传统依赖于人工统计的方式存在效率低、易出错等问题，提出基于改进深度霍夫线条先验网络(Hough Transform-Line Convolutional Neural Network，HT-LCNN)线段检测模型的隧洞施工活动时间信息智能提取方法。首先，采用单应性变换手段进行施工日志图像预处理，解决原始图像存在的倾斜、旋转、扭曲等问题;其次，利用全局上下文注意力模块(Global Context Network，GCNet)改进HT-LCNN模型的残差模块，通过建立和共享全局注意力图，获得目标线段在特征图和通道间的长距离依赖关系，提高模型对目标手绘线段的注意力，克服原有HT-LCNN方法容易受到表格线段和文字干扰的不足，实现手绘线段的高精度智能检测;进一步地，建立施工时刻-活动坐标系，根据所检测的手绘横道图线段的端点坐标位置特征，将其自动转化为施工活动时间信息。将该方法应用于某长距离引水隧洞TBM施工日志活动时间提取，本文提出的改进HT-LCNN模型的检测精度AP5、AP10、AP15值分别为94.7%、95.0%、95.1%，均高于HT-LCNN和LCNN;基于本文方法自动提取的施工活动时间与人工提取结果相比，平均绝对误差仅为1.82 min。本研究为隧洞施工活动时间信息准确高效提取提供了新思路。 |
| 英文摘要: |
| Extraction of activity duration information is important for schedule analysis and decision of tunnel construction.A common way to record the activity durations of tunnel construction is drawing the line segment as Gantt chart in the construction log.However，the traditional extraction method that relies on manual statistics is facing a dilemma between efficiency and accuracy.To solve this problem，this paper adopts a line segment detection method based on deep learning.An automatic extraction method of tunnel construction activity duration information based on improved Hough Transform-Line Convolutional Neural Network(HT-LCNN)is proposed.First，the preprocessing is based on homography to solve the obliquity，rotation and twist of the original photos.Second，the global context network(GCNet)，which has the ability of establishing the long-distance dependence of the feature graph and the dependence between channels，is integrated into the residual module of HT-LCNN to improve the model’s attention to the target line segment of Gantt chart，reduce the interference from the line segment of the table and text，and improve the detection accuracy.Third，the coordinates of detected lines are automatically converted into construction activity durations.The proposed method is used to extract the activity durations of a TBM diversion tunnel project from 17 months of construction logs.The case study shows that the detection precision AP5，AP10 and AP15 are 94[BF].[BFQ]7%，95[BF].[BFQ]0% and 95[BF].[BFQ]1% respectively，which are higher than the results of HT-LCNN and LCNN.Compared with manual extraction results，the mean absolute error of our automatic extraction method is only 1[BF].[BFQ]82 min.This study provides a novel idea for extraction of tunnel construction activity duration information accurately and efficiently. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |