城市重大基础设施灾害风险主动感知与精准管控

The goal of this project is to utilize big data to identify the damaged infrastructures in an urban infrastructure network, predict reliability of the urban major infrastructure network, figure out and give early warning on the typhoon events and earthquake events based on disaster feature extraction, and make decision on strategies of maintenance of the urban infrastructure network before disaster events and repair strategies post-disaster events (hazard risk control). To reach this goal, the active and smart sensing technology for monitoring the long-term behavior and disaster features of infrastructures will be first developed, including the mobile crowd-sensing technique, the multi-scale video monitoring technique and fast moving vehicle-parasitically distributed sensing technique. Considering the raw big data collected from different sensors may have different resolution and different accuracy, and may be unbalance-prone, the Generative Adversarial Networks (GAN) and deep learning framework with multi-layer inputs will be proposed to preprocessing the raw dataset. Based on the pre-processed big data, the graphic models for vehicle load and environmental actions over the urban infrastructure networks will be established and the spatio-temporal distribution or spatio-distribution will be investigated by using clustering algorithms. Deep learning frameworks will be proposed to mine the big data collected from long-term operation of infrastructures, agent models denoting the long-term condition of infrastructure will be established based on the features of hidden layers and output layer of the deep learning framework. Furthermore, the agent model-based graphic models will be established for the performance evaluation of urban infrastructure networks, and the critical edge, i.e. critical infrastructures in an urban infrastructure network, will be figured out from the graphic model using the big data. Considering that only very few elements in an infrastructure or few infrastructures in an urban infrastructure network will be damaged, the hierarchical sparse Bayesian learning algorithm is proposed for damage detection of infrastructure and infrastructure network according to the parameter variation of agent models. However, for the earthquake, the damaged infrastructures will be not sparsely distributed in the network, additionally, considering the data may be presented in time sequence (acceleration) or presented by spatio feature (images or video), two kinds of deep learning frameworks will be proposed for earthquake-induced damage detection of urban infrastructure network, the correlation between earthquake-induced damaged infrastructures will be further studied, and the influence of spatio-distribution of earthquake ground motion on the damaged infrastructure distribution will be also analyzed. The reliability of the urban infrastructure network and its real-time updating will be calculated by using the hierarchical sparse Bayesian learning algorithm, and the failure modes and mechanism will be studied. Combination of deep learning framework and enhancement learning framework for decision-making on the hazard risk control of infrastructure and infrastructure network will be proposed. Finally, the software code will be developed and a demo of a network consisting of bridges, viaduct and tunnels in a city will be set up based on cloud parallel computing technology on Tianhe-2. The theory for the big data-based intelligent disaster prevention and hazard risk control of urban infrastructure will be achieved, which will kick off the new era of disaster mitigation and prevention in civil engineering field.

研究基础设施灾害移动群智感知、多尺度视频监测、移动车载分布式监测等主动监测与智能感知技术；针对基础设施多源异构大数据的跨源跨尺度性、分布偏颇性、精度非一致性等问题，研究数据预处理的生成对抗网络和数据非齐次输入深度学习框架；研究车辆荷载网络和区域环境作用空间模型及空间分布特征；设计城市重大基础设施及网络运行监测大数据深度学习框架，建立运行状态深度学习代理模型与概率图；研究基础设施及网络层次稀疏贝叶斯损伤识别算法、地震破坏识别时-空域深度学习框架，揭示损伤破坏空间分布特征与机理；研究基于物理模型和代理模型的基础设施网络可靠度概率图及层次稀疏贝叶斯学习实时更新算法，揭示网络失效模式和关联机制；研究基础设施灾害风险管控深度强化学习框架；基于天和二号云超算并行技术开发集成软件并建立示范网络。本项目研究将形成基于大数据的城市重大基础设施智慧防灾与风险管控理论，开辟数据驱动土木工程防灾减灾新方向。

本项目建立了城市重大基础设施灾害风险主动感知与精准管控的系统理论与方法。提出了系列城市重大基础设施主动监测方法与智能感知技术，包括移动群智感知技术与端边云协同联邦学习机制、基于计算机视觉的感知方法、基于深度学习的结构高分辨率位移场识别与基于摄影测量与深度学习的大跨度斜拉桥三维重建方法；建立了重大基础设施车辆荷载网络时-空模型，揭示了重大基础设施网络区域环境作用时-空模型及空间分布规律；提出了基于卷积神经网络的时域-频域信息融合异常数据诊断与重构方法，提出了数据挖掘的概率分布相关性统计分析方法与状态评估方法，提出了流体力学的机器学习方法，发展了大跨桥梁抖振响应与非线性颤振动力学机器学习建模与预测方法；提出了基础设施多类型损伤识别的计算机视觉与贝叶斯网络方法，建立了结构可靠性分析的深度学习方法，提出了基础设施网络结构的图神经网络模型，研究了社区结构对重大基础设施可靠性影响和群体依赖性对重大基础设施可靠性的影响；提出了结构智能风险管控的深度强化学习框架，研究了重大基础设施受灾害攻击下的关键节点识别高效方法；开发了基于天河二号的城市重大基础设施大数据智能处理平台，开展了港珠澳大桥、南沙大桥和虎门大桥等大跨度桥梁示范应用。研究成果发表SCI期刊论文60篇（包括 Nature Communications、PNAS 等论文2 篇），会议论文11篇，大会报告12人次，申报/获批发明专利8项，软件著作权3项，主编交通部行业强制性规范1部；获国家科技进步二等奖1项、ASCE Housner奖1项、黑龙江省科学技术奖1项、中国公路学会科学技术奖1项。