基于机器学习的铁路桥墩地震损伤快速评估

在桥梁的震后抢通工作中，桥梁结构的快速损伤评估是恢复交通的关键环节。以具有代表性的铁路矩形桥墩为研究对象，通过4组拟静力试验验证有限元建模方法的合理性，并对1 000组桥墩有限元模型分别按照纵桥向和横桥向进行耐震时程分析，通过搭建BP神经网络对地震动力响应的需求结果进行拟合，构建铁路矩形桥墩震损快速评估模型，最终通过一座三跨混凝土梁桥验证该模型的适用性。研究结果表明：配筋率、配箍率、剪跨比和轴压比是影响桥墩地震损伤的四种主要因素，长宽比、混凝土和钢筋强度是影响桥墩地震损伤的三项次要因素；当发生PGA为0.32g的设计地震时，通过数值分析和神经网络模型快速评估这两种方法计算所得桥梁四个桥墩轻微损伤概率分别为96.7%、44.6%、49.1%、96.7%和95.6%、40.4%、60.9%、95.8%,中度损伤概率分别为40.1%、1.2%、1.6%、40.1%和37.4%、2.3%、6.0%、37.7%;BP神经网络算法能够有效建立构造参数与地震响应之间的联系，输出误差处于合理范围内，回归程度较好。基于BP神经网络的桥梁地震损伤评估模型具有较好的普适性，能替代部分数值仿真计算工作。

Rapid seismic damage assessment of railway piers based on machine learning

In the aftermath of an earthquake, prompt damage assessment of bridge structures is a crucial step toward restoring traffic flow. This study focuses on representative railway rectangular bridge piers, validating the reliability of the finite element modeling method through four sets of quasistatic tests. We conducted endurance time analyses on 1 000 sets of data derived from the finite element model of bridge piers in both longitudinal and transverse directions. To fit the seismic dynamic response requirements, we constructed a BP neural network and established a rapid evaluation model for assessing seismic damage to railway rectangular bridge piers. The efficacy of this model was then confirmed through its application to a three-span concrete beam bridge. Our findings suggest that the reinforcement ratio, stirrup ratio, shear span ratio, and axial compression ratio are the four main factors affecting the seismic damage of piers. Meanwhile, the aspect ratio and the strength of both concrete and steel bars emerge as secondary factors. Under a design earthquake with a PGA of 0.32g, the probabilities of minor damage to the bridge, as calculated by numerical analysis and rapid evaluation of the neural network model, are 96.7%, 44.6%, 49.1%, and 96.7%, and 95.6%, 40.4%, 60.9%, and 95.8%, respectively. The probabilities of moderate damage are 40.1%, 1.2%, 1.6%, and 40.1%, and 37.4%, 2.3%, 6.0%, and 37.7%, respectively. The BP neural network algorithm can effectively establish the relationship between structural parameters and seismic responses, producing output errors within an acceptable range and exhibiting a high degree of regression. The BP neural network-based bridge seismic damage assessment model demonstrates excellent universality and can effectively replace some numerical simulation calculations.