CAREER: Data to Models (D2M), A Domain-Guided Translation of Sensor Data to Analytical Structural Models

点击翻译按钮获取中文摘要

Despite a demonstrated potential for autonomous asset management, data-driven methods have yet failed to impact structural engineering practice significantly. A major obstacle is that the high variability inherent in civil structures causes laboratory-developed algorithms to struggle without expert guidance. This Faculty Early Career Development Program (CAREER) award supports research to address this issue. This work introduces an approach to translating raw, multi-source sensor data into functional analytical models which realistically represent the as-built structural behavior. Three data types are considered: spatial data from images and laser scanning, non-destructive evaluation data such as ground penetrating radar, and response data from accelerometers and strain gages. In addition, the methodology introduces the use of a structural ontology (a graphical map of structural component relationships) as a way to encode domain knowledge in a way a computer can understand. This creates a link between specialized sensor measurements and broader functional models, with reduced input from an engineer. It is envisioned that the results of the research would support various future applications in infrastructure assessment, including the creation of digital twins. These anchor sensor data to an analytical model of the structure, enabling practicing engineers to interpret physical behavior more effectively, run simulations, and facilitating asset management decision. The technical approach of this work addresses three open research problems: (1) model-oriented identification of structural components from spatial and non-destructive evaluation data, (2) reasoning on encoded domain knowledge to establish structural component relationships, (3) fusing response data with component information to represent as-built behavior. To address (1) and (2), the framework uniquely combines a structural engineering ontology (domain knowledge) and probabilistic graphical models (reasoning) using a machine learning framework. It begins by identifying individual structural components (e.g. beams, reinforcement) in the data, then reasoning on their relationships and functions in 3D space. Thus, the diverse and noisy information obtained from computer vision techniques (e.g., component types, geometry) can be synthesized into a candidate structure that makes contextual sense. To address (3), candidate structures are combined with response data (accelerations, strains) via system identification techniques to infer unobservable properties such as boundary conditions, material properties, and connection stiffnesses. The final output is a calibrated model capable of analysis through standard tools like finite elements. In parallel, this work integrates research products into core engineering courses. Students will interact directly with the data, learning how create structural models of real-world structures while learning about AI and its applications in civil engineering. The structural ontology will be leveraged an interactive learning guide, allowing students to explore structural relationships as they work on a problem.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.