震后快速准确的建筑应急震损评估和震前准确高效的响应时程预测，是提升防震减灾能力、控制地震灾害损失的重要手段。经典物理驱动方法已有大量研究，但在充分利用已有数据积累、保障精度的同时提升计算效率等方面，基于机器学习（特别是深度学习）的方法表现出巨大的潜力。本文采用机器学习方法，针对建筑应急震损评估和响应时程预测问题开展深入研究，主要工作内容如下：（1）针对单体建筑，提出了基于多元地震动强度指标和机器学习算法的应急震损评估方法，开展了典型案例分析，验证了方法的准确性和效率。随后，比选了对结构震损评估最为关键的地震动强度指标，并基于数据驱动思想构建了新指标。（2）针对更加多样化的建筑应急震损评估需求，考虑不同的地震动特征表征方式和神经网络算法，分别提出了基于长短时记忆神经网络和地震动时程的区域尺度建筑应急震损评估方法，以及基于卷积神经网络和地震动小波时频图的、适用于单体及区域尺度的建筑应急震损评估方法。随后，通过典型单体、区域尺度案例分析，以及和易损性分析、时程分析法的对比，验证了两类方法的准确性和计算效率。（3）针对响应时程预测中神经网络的模型架构和训练推演方法开展研究，提出了权重金字塔网络架构和具有误差自纠偏能力的两阶段训练策略，提升了神经网络的特征提取与非线性拟合能力，并有效纠正推演过程的误差累积。随后，开展了案例分析，验证了方法的精度和效率，并与多种既有网络模型进行了对比。最后，提出了数据-物理耦合驱动的结构模拟技术路线，并基于OpenSEES开展了程序实现和案例分析。（4）针对响应时程预测中数据需求问题开展研究，提出适用于小样本情形的迭代自迁移方法，包括Deep Adaptation Network with Three Branches for Regression（DAN-TR）迁移学习网络、迁移源域高效构建策略和基于伪标签策略的数据增广方法。随后，基于小样本数据集和多种网络架构开展案例分析，验证了所提出的方法的可靠性。
Efficient and accurate building seismic damage assessment under emergency situations and structural response time-history prediction before the earthquakes are important manners for improving the ability to control the loss caused by earthquakes. There have been a lot of studies focusing on classical physics-based methods, while techniques based on machine learning (especially deep learning) show great potential in making comprehensive use of existing data and improving the computational efficiency while maintaining the accuracy. With regard to the demand on seismic damage assessment and time-history prediction of buildings, following works are performed:(1) With respect to the individual building scale, an emergency seismic damage assessment method based on multiple ground motion intensity measures and machine learning algorithms is proposed, and case study is carried out to evaluate the efficiency and accuracy of the proposed method. Subsequently, the most critical ground motion intensity measures are obtained through iterative comparison, and new intensity measures are developed through data-driven approach.(2) In response to a wide variety of scenarios for emergency building seismic damage assessment, another two methods using different ground motion representation pattern and neural network algorithms are proposed. The first one is a regional-scale method based on long short-term memory neural network (LSTM) and ground motion time-series. The second one is a method based on convolutional neural network (CNN) and ground motion time-frequency distribution graphs, which can fulfill the demand on both individual building and regional scales. The accuracy and efficiency of these two methods are verified through cases on both typical building and regional scales, together with the comparison with fragility analysis and nonlinear time-history analysis.(3) In terms of the neural network architecture and training/inference strategy, the weighted pyramid neural network architecture and corrective two-stage training strategy are proposed. These methods are beneficial for improving the network models’ ability of feature extraction and nonlinear fitting, and effectively correcting the error accumulation during the inference stage. Subsequently, the accuracy and efficiency of the proposed methods are evaluated through case studies and comparisons with several commonly used neural network models. Finally, a data-physics coupling driven structural simulation framework is proposed. Based on the open-access simulation software OpenSEES, the programs are developed and case study are performed.(4) In terms of the data requirement of neural network, the iterative self-transfer learning method suitable for small datasets is proposed, including the novel network architecture “deep adaptation network with three branches for regression (DAN-TR)”, efficient construction strategy of source domain for transfer learning, and data augmentation methods based on pseudo-labelling strategy. Subsequently, case studies are carried out based on several small datasets and different network architectures to verify the reliability of the proposed method.
