2025年"核理安邦“联合博士生学术论坛

基于物理引导两阶段深度学习的建筑结构地震易损性分析

Not scheduled

12m

Location: 清华大学核能与新能源技术研究院

口头报告 AI+ AI+

Speaker

刘 雪天 (清华大学)

摘要

建筑结构地震易损性分析对于地震灾损评估具有重要意义。本研究提出了一种基于物理引导两阶段深度学习的建筑结构地震易损性分析方法，可以根据相对低成本的数据集得到建筑结构工程需求参数更精准的预测结果。这种深度学习方法将同类型建筑结构需求参数基于简化、精细模型计算结果之间的预测增量为物理引导，建立了同类型建筑结构不同物理模型之间的联系，通过两阶段集成了大规模简化模型数据集预测结果和精细模型数据集预测增量。以MLP神经网络为基础，基于本研究深度学习方法构建的神经网络验证了有效性。计算结果表明，本研究方法预测结果相较数据驱动神经网络有所改善，测试集上MSE、RMSE以及MAE等评估指标均下降，且测试集上R2达到99%所需训练集规模下降了12.98%。总体而言，所提出方法在降低地震易损性分析计算成本，实现地震情境下对破坏程度的快速、准确预测具有重要潜力和应用前景。

Abstract

Seismic fragility analysis of building structures is important for evaluation of earthquake-induced damage. This research proposes a physics informed two stage deep learning method for seismic fragility analysis of building structures, which provides better prediction of engineering demand parameters with less data cost. The proposed deep learning method utilizes the increment of predictions of engineering demand parameters between simplified and refined model under the same seismic scenarios as physics information, establishes connections between different physical models of the same structural type, effectively integrating predictions from large-scale simplified model datasets and the increment of predictions from refined model datasets with two stage techniques. The neural network developed based on the multilayer perceptron and implemented within the proposed deep learning method demonstrates the effectiveness of the proposed methodology. The computational results demonstrate that the proposed method achieves significant improvements over conventional neural networks. Notably, evaluation metrics on the test set show substantial reductions. Furthermore, the method reduces the required training dataset size by 12.98% to achieve an R² value of 99% on the test set. Overall, the proposed method exhibits considerable potential in reducing computational costs associated with seismic fragility analysis, while facilitating rapid and accurate prediction of damage levels under various seismic scenarios.