基于深度学习的钢结构锈蚀程度定量评估方法OA
Quantitative Evaluation Method for Corrosion Degree of Steel Structures Based on Deep Learning
针对现有钢结构锈蚀评估采用经验分级、难以有效评估结构力学性能的问题,提出一种基于深度学习的定量评估方法.制备不同锈蚀程度的钢板试件,采用3D激光扫描获取锈蚀表面形貌,分别通过体积积分得到体积损失率(DOV)和通过有限元逆向建模得到锈后最大应力,以两者作为图像的量化标签构建锈蚀图像的数据集,采用改进的ResNet-18模型预测DOV,采用预训练与微调后的ViT-B/16模型预测锈后最大应力.结果表明:改进的ResNet-18模型对DOV的预测误差控制在0.015以内;ViT-B/16模型训练稳定后拟合度约为0.65,验证损失稳定在0.01~0.02.表明该方法可实现锈蚀程度的连续定量评估,为钢结构锈蚀检测与运维决策提供了新的技术路径.
Aiming at the drawbacks of empirical classification adopted in the existing corrosion evaluation of steel structures,which fails to effectively assess the mechanical properties of structures,a quantitative evaluation method based on deep learning is proposed in this paper.Steel plate specimens with various corrosion degrees are fabricated,and a 3D laser scanner is utilized to acquire the surface topography of corroded specimens.The degree of volume loss(DOV)is calculated via volume integration,while the maximum post-corrosion stress is obtained through inverse finite element modeling.These two indices are taken as quantitative labels of corrosion images to establish a corrosion image dataset.An improved ResNet-18 model is adopted to predict DOV,and a pre-trained and fine-tuned ViT-B/16 model is employed to forecast the maximum post-corrosion stress.The experimental results demonstrate that the predic-tion error of DOV by the improved ResNet-18 model is controlled within 0.015.After stable training of the ViT-B/16 model,the coefficient of determination reaches approximately 0.65,and the validation loss remains steady at 0.01~0.02.It is verified that the proposed method can realize continuous quantitative evaluation of corrosion severity,offer-ing a novel technical route for corrosion inspection and operation-maintenance decision-making of steel structures.
张列学
上海市道路运输事业发展中心,上海 200025
建筑与水利
钢结构锈蚀程度定量评估深度学习体积损失率锈后最大应力
steel structurequantitative evaluation for corrosion degreedeep learningdegree of volume loss(DOV)maximum post-corrosion stress
《市政技术》 2026 (8)
111-118,8
评论