| Author: | Hao, Shuo |
| Title: | Advanced machine learning approaches for force reconstruction and state estimation in structural dynamics |
| Advisors: | Ni, Yi-qing (CEE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Civil and Environmental Engineering |
| Pages: | xxii, 221 pages : color illustrations |
| Language: | English |
| Abstract: | Forward problems in the structural dynamics typically involve simulating structural responses under specified conditions through constructed structural models, which are fundamental problems from the design of new structures and assessment of existing structures. Conversely, inverse problems in structural engineering often serve as dual counterparts to many forward problems, with their models typically evolving from various forward problem models. In existing structures, modern sensing technologies can effectively acquire structural information, enabling the inference of valuable structural insights based on measured data and inverse problem models. Generally, valuable structural information refers to structural parameters, external forces acting on the structures, and responses at unmeasured locations of structures. The first part of the thesis introduces a nonparametric Bayesian multi-task learning framework for time-domain force reconstruction. Traditional methods often struggle with ill-posed deconvolution problems and uncertainties. By assigning the Gaussian process (GP) priors to force functions within a Bayesian context, the proposed method effectively mitigates these issues. This approach leverages the relationship between loads and responses through the convolution operator, resulting in responses that also follow a GP. A joint Gaussian distribution across multiple tasks enables closed-form posterior distributions of the forces based on measured responses. The framework is validated through simulations on a truss bridge and experiments on a frame structure subjected to impact loads, demonstrating high accuracy and efficient uncertainty quantification. Building on this framework, the thesis extends its application to reconstruct transient aerodynamic loads on high-speed maglev vehicles using onboard acceleration measurements. Traditional estimation methods are often time-consuming and costly. The novel framework utilizes an inverse mathematical model derived from a calibrated maglev vehicle model and integrates the multi-task GP algorithm to reconstruct aerodynamic loads efficiently. The method treats the reconstructed loads as multiple GP, allowing for closed-form posterior calculations. Validation with data from maglev trains passing through double-track tunnels confirms the framework's effectiveness. To enhance state estimation capabilities, the thesis presents two complementary approaches. The first is a Bayesian-based multi-fidelity GP method that employs a time-delayed GP model to capture the statistical correlations between sensor observations from preceding time steps and the current structural states. Training data for this model is generated using a simulated finite element model subjected to synthetic excitations. To further enhance the accuracy, the proposed multi-fidelity GP model integrates data from both simulations and actual measurements. This integration effectively combines the low-fidelity information from simulations with high-fidelity direct measurements to provide posterior estimations with quantified uncertainties. Validation examples confirm the effectiveness of multi-fidelity GP in providing precise state estimations, thereby expectedly boosting the prediction reliability across various applications in structural dynamics. The second approach explores a recurrent neural network (RNN) method enhanced with transfer learning. While GP offer probabilistic estimations, the RNN provides high-accuracy point estimations by learning correlations in multi-output problems. The RNN is trained using extensive response data from a calibrated finite element model under synthetic excitations. Transfer learning adapts the RNN to real-structure predictions using actual measurement data. A fine-tuning strategy adjusts parameters in the RNN initial layers while keeping output layers fixed, ensuring effective convergence. Numerical and experimental validations show that the RNN models significantly outperform traditional methods, especially in complex, multi-output state predictions. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13792