北京工业职业技术学院学报

随着中国隧道工程规模不断扩大，衬砌表观病害对隧道安全运营构成严重威胁。传统检测方法在复杂环境下存在局限性，而深度学习技术为病害识别提供了新途径。采用VGG模型，针对裂缝、剥落和渗水3类常见隧道衬砌病害，构建了包含2 700张图像的数据集。对比分析了VGG16模型和引入注意力机制的VGG16-SE模型，实验结果表明：经过100轮训练后，VGG16-SE模型准确率达到92.4%,损失值6.5%;与VGG16模型相比，准确率提升11.6%,损失值降低53.9%。VGG16-SE模型能更有效地识别隧道衬砌表观病害，为混凝土结构病害的智能检测提供参考。

With the continuous expansion of tunnel construction scale in China, tunnel lining surface defects pose a serious threat to the safe operation of tunnels. Traditional detection methods have limitations in complex environments, while deep learning technology provides a new approach for defect identification. In this study, a dataset containing 2 700 images was constructed using the VGG convolutional neural network to identify three common tunnel lining defects: cracks, spalling, and seepage. Two models, VGG16 and VGG16-SE with an attention mechanism, were compared and analyzed. The experimental results show that after 100 training epochs, the accuracy of the VGG16-SE model reached 92.4%, with a loss rate of 6.5%. Compared with the VGG16 model, the accuracy increased by 11.6% and the loss rate decreased by 53.9%. The research indicates that the VGG16-SE model can more effectively identify tunnel lining surface defects, providing a reference for the intelligent detection of concrete structure defects.