Physics-Informed Data-Driven Methods for Long-Term Structural Health Monitoring of Concrete Structures

Abstract

Concrete exhibits long-term time-dependent behavior due to rheological effects that impacts the safety and serviceability of civil infrastructure, such as bridges and high-rise buildings. In real concrete structures, long-term prediction is challenging due to the presence of variable environment conditions, such as temperature, humidity, as well as uncertainty in loading and boundary conditions, and the random nature of creep and shrinkage. Existing cutting-edge methods for prediction of long-term time-dependent behavior at structural scale involves computationally intensive numerical methods in which several semi-empirical rheological models are applied. However, these models are targeted at structural design and adjusted based on experimental data that may not be the most accurate for an existing structure. Structural health monitoring can improve long-term prediction by providing structure-specific in-situ measurements. This dissertation introduces novel methods for the prediction of long-term behavior of concrete structures that integrate structural health monitoring, structural modeling and analysis, and machine learning. A new method proposed integrates probabilistic neural networks and analytical structural model, together with generalized creep and shrinkage models, for the detection of gradual anomalies in real structural data that are difficult to detect with existing methods. A creative integration of analytical structural modeling and neural networks is used for reconstruction of 2D normal strain field on a real structure over multiple years. Finally, a new method for the prediction of long-term behavior in high-rise buildings integrates analytical modeling with generalized creep and shrinkage models, with rigorous uncertainty quantification, is presented. The dissertation contributes to the literature in long-term monitoring of structures with innovative integration of structural modeling, and machine learning.