我国已建成大量的高拱坝。这些高拱坝产生了巨大的经济社会效益，但失事造成的后果是不可接受的。有效的高拱坝健康监测和预警分析是保障拱坝安全运行的重要前提。由于高拱坝运行条件复杂、监测数据稀疏、非线性强等难点的存在，传统的单测点行为模型和预警指标在高拱坝安全监测的时效和准确性方面存在局限性。针对以上的研究难点，本论文融合深度学习、先验物理知识、卫星遥感以及有限元模型等多种方法，研究了高拱坝运行期的变形时间效应、温度场、变形场以及监测预警指标，并基于实际大坝运行性态进行模型和方法的验证。本文的主要研究内容和创新成果如下：（1）综合考虑温度非线性、滞后效应和时序特性，建立了基于长短期记忆神经网络的高拱坝单测点变形行为模型，该模型相比传统的HST、HSTT模型以及MLP神经网络模型具有更高精度，且具有更强的处理高拱坝长期安全监测数据短期异常、长期时效的能力。（2）考虑卫星遥感天气数据、热传导基本原理，提出了基于卷积神经网络的高拱坝下游面温度场预测模型，实现了小湾拱坝下游面温度场的精确重建，结果表明拱坝下游面温度场具有显著的不均匀分布特征，且受太阳辐射、天气条件影响。采用小湾拱坝确定的模型参数，无需大坝实测数据，预测了高精度的二滩拱坝下游面运行期温度场，表明该模型具有很好的可迁移性，可用于预测在建或未建拱坝的下游面运行期温度场，作为拱坝设计温度荷载的补充。（3）提出新的基于变形场的高拱坝结构开裂预警指标，采用有限元预设多种不同的开裂和劣化情况计算高拱坝运行变形数据，以此验证了预警指标的有效性和敏感度，结果表明，新的预警指标可获取拱坝开裂发生时刻，并精确识别开裂区域。（4）考虑高拱坝结构和荷载特性，结合本文所提出的基于卷积神经网络的高拱坝下游面温度场预测模型获取的真实温度场，建立了基于卷积神经网络的高拱坝全局统一的变形场预测模型。该模型的变形场预测精度高于经过精细反馈分析的有限元模型，结合变形场开裂预警指标，实现了高拱坝长期运行健康监控。

A large number of high arch dams have been constructed in China. These high arch dams have created huge economic and social benefits while the consequences of dam failure are serious and unacceptable. Effective health monitoring and early warning analysis of high arch dams are necessary to ensure the safe operation of arch dams. However, due to the complicated operating conditions, sparse monitoring data, strong nonlinearity behaviors and other research difficulties of high arch dams, the traditional behavior model and early warning index based on single monitoring point is limited by the timeliness and accuracy for high arch dam safety monitoring. Aiming at the above research difficulties, this paper studies the time-varying effect of deformation, temperature field, deformation field, and early warning index of high arch dams during their operation period by integrating various methods, such as deep learning, prior physical knowledge, satellite remote sensing and finite element model, and verifies the proposed models and methods based on the real behavior of actual dams. The main works and innovations of this research are as follows:(1) Considering nonlinearity, delay effect of temperature and time series characteristics, the single point deformation behavior coupling model of high arch dams based on a long short-term memory neural network is established. Compared with the traditional HST, HSTT model and MLP model, coupling model has higher accuracy and a stronger ability to deal with short-term anomalies and separate real long-term time-varying of high arch dams.(2) Considering satellite remote sensing weather data and the basic principle of heat conduction, a predictive model of the temperature field on the downstream face of the high arch dam based on convolutional neural network is proposed, which accomplished an accurate reconstruction of the temperature field on the downstream face of the Xiaowan arch dam. The results show that the temperature field on the downstream face of the arch dam is obviously uneven and is significantly affected by solar radiation and climatological conditions. Using the model parameters determined with the Xiaowan arch dam, the temperature field of the downstream face of the Ertan arch dam in the operation period is predicted accurately without any monitoring data of the dam, which shows that the model has the enormous generalization ability of predicting the downstream face temperature field of arch dam under construction or designation, and thus it could be used as a supplement to the design temperature load of the arch dam.(3) New early warning indexes of high arch dam cracking based on the deformation field are proposed. The effectiveness and sensitivity of the early warning index are verified through the deformation field data calculated with a finite element model, which preset various cracking and deterioration conditions. The results show that the new early warning indexes can quickly obtain the cracking time and accurately identify the cracking area of the arch dam.(4) Considering the structure and load characteristics of a high arch dam, combined with the real temperature field, a global unified deformation field prediction model of a high arch dam is established based on a convolutional neural network. The prediction accuracy of this model for the deformation field is higher than that of the finite element model after detailed feedback analysis. Combined with the early warning index of deformation field cracking, long-term operation health monitoring of high arch dams is realized.