面向结构地震响应预测的Phy-LInformers方法

扫码查看

原文链接

万方数据
维普

中文摘要：为了准确评估建筑结构在地震作用下的动力特性和延性性能并促进韧性城乡的建设,本文提出了一种名为Phy-LInformers的深度学习框架,该框架综合运用了长短期记忆网络(long short-term memory,LSTM)、Transformer类模型Informer以及物理先验知识,以实现对建筑结构非线性地震响应的精确预测.该框架的核心思想是结合Informer的编码(Encoder)和解码(Decoder)结构,在Decoder部分引入了LSTM以预测建筑物先前的历史状态信息.同时,通过将现有的物理知识(例如预测变量之间的状态依赖关系和运动控制方程等)编码到损失函数中,对Phy-LInformers进行指导并约束其学习空间,同时提高有限训练数据下深度学习模型的预测性能.随后,通过 2 个模拟数据算例验证所提框架的性能.结果表明,所提出的Phy-LInformers是一种鲁棒性良好、预测性能优秀的非线性地震响应预测方法,即使在训练样本非常少(例如仅有 10 条)的情况下依然能准确预测结构在地震作用下的动力响应.这一特性使得Phy-LInformers在工程实践中具有可行性,并且在建筑结构抗震性能评价领域展现出良好的应用前景.

外文标题：Phy-LInformers approach toward structural seismic response prediction

外文摘要：In order to accurately assess the dynamic and ductile properties of building structures under seismic action and to promote the construction of resilient cities and towns,in this study,we introduce a novel deep learning frame-work denoted as Phy-LInformers,which integrates long short-term memory(LSTM),Transformer-based model Inform-er,and prior physical knowledge to achieve precise prediction of nonlinear seismic responses in building structures.The central concept of Phy-LInformers lies in fusing the Encoder and Decoder architectures of Informer,while integrating LSTM within the Decoder component to forecast preceding historical states of the building.Meanwhile,the learning space of the Phy-LInformers'training process is instructed and constrained by encoding existing physical knowledge(e.g.,state dependencies between predictor variables and motion control equations,etc.)into the loss function.And at the same time,the prediction performance of the deep learning model is improved with limited training data.Sub-sequently,the satisfactory performance of the proposed framework is successfully demonstrated through two illustrative examples.The results demonstrate that the proposed Phy-LInformers is a nonlinear seismic response prediction method with better robustness and superior prediction performance,which can still accurately predict the dynamic response of a structure under seismic excitation even with very few training samples(e.g.,only 10 samples).This feature makes Phy-LInformers feasible for engineering practice and shows promising application prospects in the field of seismic perform-ance evaluation of building structures.

外文关键词：

seismic response predictionphysical knowledgephysics-informed deep learningtime series forecastingfew shot learningInformerlong short-term memoryPhy-LInformers

作者：

郭茂祖、张欣欣、赵玲玲、张庆宇

展开 >

作者单位：

北京建筑大学电气与信息工程学院,北京 100044

北京建筑大学建筑大数据智能处理方法研究北京市重点实验室,北京 100044

哈尔滨工业大学计算学部,黑龙江哈尔滨 150001

关键词：

地震响应预测物理知识物理驱动的深度学习时间序列预测少样本学习 Informer 长短期记忆网络 Phy-LInformers

基金：

国家自然科学基金项目北京市自然科学基金项目

项目编号：

622710364232021

出版年：

2024

DOI：

10.11992/tis.202404002

智能系统学报

中国人工智能学会　哈尔滨工程大学

智能系统学报

CSTPCD北大核心

影响因子：0.672

ISSN：1673-4785

年,卷(期)：2024.19(4)

参考文献量7