基于SMR-RCNN网络的隧道衬砌缺陷智能识别研究

Research on Intelligent Recognition of Tunnel Lining Defects Based on SMR-RCNN Network

脱空和不密实是隧道衬砌最常见的两种病害。在这两种病害长期作用下会导致隧道出现破裂、渗漏水、钢筋锈蚀,最终造成隧道塌方等问题,严重威胁行车安全。采用探地雷达对隧道进行无损探测是发现这些病害或缺陷的常见方式,但大量雷达数据的人工识别存在着工作量大、效率低、强烈依赖人员的专业素养等问题。本文提出一种基于深度学习的隧道衬砌缺陷的自动检测方法——自监督多尺度池化区域卷积神经网络方法(Self-monitoring Multi-scale ROI Align Region Convolutional Neural Network,SMR-RCNN),以提高缺陷识别的效率,并减少主观因素的影响。在雷达探测隧道衬砌的实践中,数据量巨大,但缺陷样本却很少,这对训练神经网络是一个相当大的挑战。为此,设计了一种数据增强的方法来增加缺陷的样本数量,且使用一种自监督对比学习的网络模型来提取雷达数据的特征,然后将其迁移到改进后的Faster-RCNN网络模型中;最后,使用有标签的样本对改进的Faster-RCNN网络进行细调训练。实验结果表明,相较于传统的Faster-RCNN方法,本文提出的算法增强了神经网络对脱空和不密实两类缺陷的自动识别能力,在检测精度上得到了显著提高,mAP值提升了12%。

The voiding and incompactness are the two most common types of tunnel linings.Their long-term impacts can lead to tunnel cracks,water leakage,corrosion of rebars,and ultimately tunnel collapse,posing a seriously threat to traffic safety.To detect these defects,ground-penetrating radar(GPR),a non-destructive testing method,is commonly used.However,manual identification of large collected radar data has programs which are heavy workload,time-consuming,and strong dependence on the professionalism of personnel.In this paper,we propose a deep learning-based automatic detection method for tunnel lining defects called self-monitoring multi-scale ROI align RCNN(SMR-RCNN),which aims to improve the efficiency of defect recognition and reduce the influence of the subjective factor.In radar detection of tunnel lining,when the data volume is huge defect and samples are scarce,it poses a considerable challenge for training the neural network.To address this issue,we designed a data augmentation method to increase the number of defect samples.We utilized DenseCL network model,a self-supervised contrastive learning network model,to extract features from the radar dataset,which were then transferred to an improved Faster-RCNN network model.Finally,we fine-tuned it by using the labelled samples.Experimental results show that compared with the traditional Faster-RCNN method,the SMR-RCNN algorithm we proposed enhances the neural network’s ability to automatically identify voiding and incompactness defects,significantly boosting the detection accuracy and a 12%increase in the mean average precision(mAP)value.