基于全场振动视频监测的结构状态与荷载贝叶斯识别方法

The spatial-resolution of structural sensing tools and robustness of structural identification are the major bottlenecks to realize the goal of structural health monitoring (SHM). This project will research on the digital video camera based full-field sensing and monitoring technique (every pixel measurement on the structure corresponds to a “sensor”), motivated by the fact that traditional physically-attached sensors in SHM only provide sparse, discrete point-wise measurements, yielding low spatial sensing resolution that usually induces the ill-posed and ill-conditional problems for modal-based damage detection and model updating of larger-scale structures. Associated with the video-based full-field monitoring, the high-dimensional models and unavoidable modeling errors should be treated explicitly. By introducing the theories of machine learning and Bayesian probability theories, we will investigate the automatic learning and posterior uncertainty quantification of model parameters in high-dimensional space. With regard to the research contents, firstly, we will study video monitoring of structural vibration and structural motion extraction with high spatial resolution. Then we will propose Bayesian identification methods of structural condition and loads from video-based full-field vibration monitoring. We will construct the hierarchical Bayesian learning models and then investigate how to calculate the posterior distributions of structural modal, stiffness reductions and dynamic loads efficiently. In addition, the confidence intervals of these identification results also can be estimated from their posterior uncertainties. After that, Bayesian information fusion method of distant video monitoring and close-range video monitoring will be investigated, motivated by the fact that the distant monitoring can get full-field structural information but the accuracy is poor, while the close-range monitoring works with high accuracy but it only collects the local vibration information of the structure. Finally, the structural model test and prototype monitoring will be performed for validation. The research will provide systematic statistical inversion algorithms for structural health assessment based on full-field, high-spatial-resolution video monitoring. They have both the vital scientific significance and application values.

结构传感手段的空间分辨率和识别方法的鲁棒性是实现健康监测目标的重要瓶颈。针对传统接触式传感器空间分布稀疏导致的结构损伤识别和模型更新的不适定和病态问题，本项目研究结构全场视频监测技术（每一像素对应一个“传感器”）。针对视频监测数据高维和模型误差，以机器学习和贝叶斯推理为理论基础，研究高维空间下模型参数的自学习及其后验不确定性的定量评价。首先，研究结构振动视频监测和高分辨位移信息的提取方法；其次，研究基于全场视频监测的贝叶斯结构状态和荷载识别方法，构造层次稀疏贝叶斯学习模型，计算结构全场模态、刚度和荷载参数的后验概率分布，并估计识别结果的置信区间；然后，针对大型结构远景监测精度较差、近景精度高但只能拍摄结构局部振动信息的问题，研究远景和近景监测的贝叶斯信息融合方法。最后，通过模型试验和原型监测进行验证。研究将形成系统的基于全场高分辨视频监测的健康诊断统计反演算法，具有重要科学意义和实用价值。

结构健康监测的传感空间分辨率不足是实现结构健康诊断目标的重要瓶颈，而全场高分辨率的结构传感手段是缓解结构健康监测反问题的不适定和病态性的重要途径。针对结构健康监测中传统接触式传感器空间稀疏布置导致结构损伤识别和模型更新的不适定和病态反演问题，本项目研究了结构振动全场视频监测技术（每一像素对应一个“传感器”）。进一步针对视频监测中视觉数据高维和反演模型误差问题，以机器学习和贝叶斯推理为手段，研究了高维空间下稀疏贝叶斯学习及模型参数的后验不确定性量化理论，实现了高空间分辨率结构监测传感和高鲁棒性诊断识别的目标。首先，研究了结构振动视频监测和高分辨位移信息的提取，建立了一种非接触式、低成本的基于深度神经网络模型的结构振动位移识别方法。通过引入循环神经网络光流估计算法来估计振动视频中的全场光流，进一步建立基于全场光流估计的结构振动位移计算方法。其次，针对视频监测的数据高维和模型误差问题，研究了高维空间下模型参数的稀疏贝叶斯学习算法框架。然后，基于结构全场空间高分辨率视频监测提取的结构高分辨率振动位移场，提出结构状态识别和荷载识别方法框架，分别建立了结构模态参数的自动识别方法、基于层次稀疏贝叶斯学习的结构损伤评估方法、基于计算机视觉的桥梁车辆荷载识别方法以及基于动态稀疏系统贝叶斯学习的结构荷载和刚度演化识别方法；最后，针对大型结构远景监测精度较差、近景精度高但只能拍摄结构局部振动信息的问题，研究了基于远景监测和近景监测的多摄像机空间坐标一致化和贝叶斯信息融合方法，提高了视频监测方法的实用性。以上研究成果通过多个结构模型试验和原型监测数据进行了验证，并通过对比试验验证了本项目方法良好的准确性和实用性。本项目研究成果形成了系统的基于全场高分辨视频监测的健康诊断统计反演算法，丰富和发展了结构健康监测理论，具有重要科学意义和实用价值。