摘要
隧道光面爆破设计时,往往需要光爆效果作为设计依据,以实现爆破安全施工并提高爆破效率。针对目前光爆眼痕识别过程中存在的现场环境复杂、检测困难等问题,提出基于DMT-U^(2)-Net与self-attention模块的复合算法模型进行爆破眼痕识别。采集爆破工程中常见的爆破眼痕图像样本,并对数据进行增强、三维重建与降噪处理,构建DMT-U^(2)-Net网络模型并改进损失函数对眼痕图像进行训练,获取DMT-U^(2)-Net眼痕分割模型;将DMT-U^(2)-Net模型处理后的分割图片与三维重建模型进行特征融合,构建基于self-attention模块的回归预测模型对融合特征进行训练,获取眼痕长度回归预测模型;将DMT-U^(2)-Net眼痕分割模型与基准U^(2)-Net,U-Net,DeepLab v3,FCN,LR-ASPP网络模型的眼痕分割结果进行对比,从而评估其训练效果;将回归预测模型与bp,GRU模型进行对比,并对输入参数进行敏感性分析,优化网络参数输入并评估网络训练效果。结果表明,DMT-U^(2)-Net网络模型分割可见眼痕的P_(DSC),P_(pre),P_(rec),P_(mIOU)分别为90.89%,91.11%,91.01%,91.59%,模型大小仅为19.76 MB,相较基准模型缩减88.2%。与其他模型相比,该模型在分割精度和模型大小,都具有较大优势;通过回归预测模型,可以实现对可见眼痕长度的精准预测,模型决定性系数高达0.992,模型大小仅为154.1 KB。将本文复合算法模型应用于隧道光面爆破可见眼痕的识别中,模型展现出较好的识别效果,基本实现了可见眼痕的端到端识别,为隧道的超欠挖识别与智能评价系统打下坚实基础。
In tunnel smooth blasting design,the smooth blasting effect is often used as a design basis to ensure safe blasting operations and improve blasting efficiency.To address the challenge in identifying smooth blasting traces,such as complex field environments and detection difficulties,this paper proposed a composite algorithm model for blasting trace recognition based on the DMT-U^(2)-Net and self-attention modules.First,common blasting trace image samples from blasting projects were collected and processed through data enhancement,3D reconstruction,and denoising.A DMT-U^(2)-Net network model was constructed,and the loss function was improved to train the trace images,resulting in the DMT-U^(2)-Net trace segmentation model.The segmented images produced by the DMT-U^(2)-Net model were fused with 3D reconstruction model features to build a regression prediction model based on the self-attention module.This regression model was trained on the fused features to predict trace lengths,resulting in a trace length regression prediction model.The DMT-U^(2)-Net trace segmentation model was compared with baseline models,including U^(2)-Net,U-Net,DeepLab v3,FCN,and LRASPP,to evaluate its segmentation performance.The regression prediction model was further compared with BP and GRU models,and sensitivity analysis of input parameters was conducted to optimize network parameter inputs and assess network training performance.The results show that the DMT-U^(2)-Net network model achieves segmentation accuracies of P_(DSC),P_(pre),P_(rec),and,P_(mIoU),at 90.89%,91.11%,91.01%,and 91.59%,respectively.For visible trace segmentation,with a model size of only 19.76 MB,representing an 88.2%reduction compared to the baseline model.Compared to other models,this model demonstrates significant advantages in both segmentation accuracy and model size.Moreover,the regression prediction model can achieve precise predictions of visible trace lengths,with a coefficient of determination as high as 0.992 and a model size of only 154.1 KB.Applying this composite algorithm model to the recognition of visible traces in tunnel smooth blasting can demonstrate excellent recognition performance.This study can effectively achieve end-to-end recognition of visible traces,laying a solid foundation for the identification of over-excavation and underexcavation in tunnels and the development of intelligent evaluation systems.
作者
凌同华
谢长庚
曹峰
廖逸轩
袁宇
LING Tonghua;XIE Changgeng;CAO Feng;LIAO Yixuan;YUAN Yu(School of Civil Engineering,Changsha University of Science and Technology,Changsha 410014,China;Hunan Provincial Communications Planning,Survey&Design Institute Company Limited,Changsha 410008,China)
出处
《铁道科学与工程学报》
北大核心
2025年第9期4248-4259,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(52478386)。
