Degree
Doctor of Philosophy (PhD)
Department
Department of Civil and Environmental Engineering
Document Type
Dissertation
Abstract
Accurate response prediction and system identification are critical for understanding complex dynamic systems. These capabilities enable effective structural design, control, and monitoring, ensuring structural integrity and safety. Despite significant progress that has been achieved in the past decades, several key challenges still remain: (1) discontinuities (e.g., joint connections) in complex structures are often represented by simplified models, which can fail to capture the real structural behavior and result in inaccurate damage identification.; (2) structural damage induces nonlinear and time-varying dynamics that traditional methods struggle to predict; (3) inevitable modeling errors degrade response prediction accuracy; (4) environmental variability and nonstationary noises complicate system state estimation; and (5) the high dimensionality of complex systems hinders model training and compromises system learning accuracy.
To address these issues, this study develops hybrid model-based and data-driven approaches integrating physics-informed neural networks (PINNs) with enhanced algorithms for efficient response prediction and damage identification. First, to identify damage in structures with discontinuities, such as joint loosening in pipeline systems, a novel joint element model is proposed to simulate the mechanical behavior of loosened joints in pipeline systems, offering improved accuracy over reduced-stiffness methods. Second, a real-time updated physics-informed neural network (RTU-PINN) is proposed, enhancing recurrent neural networks with Runge-Kutta schemes and neural ordinary differential equations (neural ODEs) to capture nonlinear, time-varying dynamics. Third, a physics-guided variational autoencoder with a variational Bayesian adaptive Kalman filter (PGVAE-VBAKF) is developed to address modeling errors and time-varying noises, and a residual neural network-based reduced-order strategy is adopted to accelerate training and improve the learning accuracy for high-dimensional systems. Finally, an unsupervised Deblurring-Deep Video Stabilization (DB-DVS) method is proposed for vision-based 3D response sensing, effectively addressing camera motion disturbances.
Extensive simulations and experiments validate the effectiveness of these methods. Results demonstrate accurate identification of structural damage, robust prediction of nonlinear responses, reliable uncertainty quantification under nonstationary noise, and improved stability of vision-based monitoring under camera motion. Overall, the proposed framework offers promising applications in structural health monitoring and digital twinning of complex engineering systems.
Date
11-12-2025
Recommended Citation
Li, Huaguan, "Complex System Response Prediction and Damage Identification via Integrating Models, Data, and Physics-Informed Neural Networks" (2025). LSU Doctoral Dissertations. 6925.
https://repository.lsu.edu/gradschool_dissertations/6925
Committee Chair
Sun, Chao