Abstract

The need for real-time, accurate monitoring of infrastructure is becoming increasingly urgent as aging structures and increasing loads put more stress on critical components. Traditional methods for structural health monitoring (SHM), which are either purely data-driven or solely based on physics, often face shortcomings. To address the limitations, this thesis investigates the use of Physics-Informed Neural Networks (PINNs), which combines data-driven approaches with physics-based principles. As part of a larger study, this thesis explores the practical application and effectiveness of PINNs in modeling the deflection of reinforced concrete (RC) beams in real world scenarios. The primary objective of the study is to evaluate PINNs performance considering complexities of a real RC beams in order to have a better insight regarding PINNs application in the real-world structural health monitoring and management as a modeling tool. To address this question, the research begins with the development and implementation of a PINN-based application for elastic beam models using numerical investigations with Finite Element Method (FEM) data for training. This initial phase establishes a foundation for integrating PINNs into structural modeling by addressing point loads and beam behavior under various conditions. Following this, the study extends the application to RC beams, incorporating practical considerations such as crack formation, reinforcement effects, and nonlinear load responses. Key contributions include the development of a practical PINN-based tool for structural deflection prediction, handling of real-world peculiarities in beam behavior, and the integration of point load effects. The thesis also investigates the impact of different loss weighting systems on model performance, providing insights into role of data and physics in training PINN in different stages of loading. Comparative analyses with FEM models highlight the strengths and limitations of PINNs, demonstrating their potential to complement traditional methods by capturing complex real-world phenomena such as crack, reinforcement interface, and nonlinearity.
Overall, the research advances the understanding of PINNs in structural engineering, showing their potential to enhance modeling accuracy and reliability in practical applications. The findings offer a valuable framework for future research and applications in structural health monitoring, paving the way for more effective and adaptable tools in the field.