{"id":{"repo_id":"exeter","oai_identifier":"oai:figshare.com:article/32603244"},"canonical_url":"https://search.dev.ndltd.org/etd/exeter/oai:figshare.com:article/32603244","repository":{"repo_id":"exeter","name":"University of Exeter","base_url":"https://api.figshare.com/v2/oai"},"display":{"title":"Physics-Informed Neural Networks for Structural Health Monitoring of Bridges","abstract":"Bridges are critical components of a nation's transport infrastructure. Structural health monitoring (SHM) is increasingly employed, particularly on key bridges, to track their structural performance and for their effective structural management. Interpretation of sensing data from SHM however remains a challenge. Conventional data-driven models often require large volumes of measurements and lack physical interpretability, while purely physics-based models can be difficult to update with sparse, noisy monitoring data. This thesis investigates physics-informed neural network (PINN)-based frameworks for interpreting bridge monitoring data, with the overarching aim of enabling physically consistent prediction of structural response and identification of key stiffness parameters using limited measurements. As PINNs are a relatively new development with almost no prior application to SHM at the commencement of this study, this research's approach has been incremental, starting from simple scenarios, and gradually ramping up complexity, examining both forward and inverse problems. A general PINN methodology is first formulated for both forward and inverse problems, including the construction of physics-based loss functions, adaptive loss weighting, collocation points sampling strategies and error metrics for model evaluation. This framework is then applied to three classes of structural systems. For a Kirchhoff–Love plate under quasi-static loading, PINN configurations with varying combinations of data, boundary conditions and governing partial differential equations are systematically compared. The results show that enforcing force and displacement boundary conditions, supplemented by sparse deflection sensors, can reconstruct plate deflections and bending moments with low bias, while explicit PDE terms further improve robustness to noise and modelling imperfections. For simply supported beams under moving loads, forward PINNs are developed using spatial–temporal inputs, and then extended to include axle load magnitude as an additional input. These models accurately reproduce deflection, moments and strains for loads both within and beyond the training range and demonstrate virtual sensing capabilities at locations that there are no physically sensors instrumented or it is not easy to measure the structural response. Inverse PINNs are then used to estimate flexural rigidity from synthetic and field strain data, achieving flexural-rigidity errors typically below 1% for noisy synthetic data and reducing the error from about 14% to less than 2% when real bridge measurements from a single sensor are incorporated. Finally, a multi-PINN framework for integrating PINNs based on domain decomposition is introduced in which a beam is partitioned into subdomains, each represented by its own network and stiffness parameter, with compatibility and equilibrium enforced at interfaces. Forward models predict full-field responses for beams discretised into multiple segments, while inverse models identify subdomain level stiffness reductions with relative errors generally below 3%, successfully localising both single- and multi-domain stiffness losses. Overall, the thesis demonstrates that PINNs provide a physically interpretable approach for bridge SHM, capable of fusing sparse monitoring data with governing physics to support virtual sensing and subdomain level system identification under moving loads.<p></p>","abstract_html":"Bridges are critical components of a nation&#x27;s transport infrastructure. Structural health monitoring (SHM) is increasingly employed, particularly on key bridges, to track their structural performance and for their effective structural management. Interpretation of sensing data from SHM however remains a challenge. Conventional data-driven models often require large volumes of measurements and lack physical interpretability, while purely physics-based models can be difficult to update with sparse, noisy monitoring data. This thesis investigates physics-informed neural network (PINN)-based frameworks for interpreting bridge monitoring data, with the overarching aim of enabling physically consistent prediction of structural response and identification of key stiffness parameters using limited measurements. As PINNs are a relatively new development with almost no prior application to SHM at the commencement of this study, this research&#x27;s approach has been incremental, starting from simple scenarios, and gradually ramping up complexity, examining both forward and inverse problems. A general PINN methodology is first formulated for both forward and inverse problems, including the construction of physics-based loss functions, adaptive loss weighting, collocation points sampling strategies and error metrics for model evaluation. This framework is then applied to three classes of structural systems. For a Kirchhoff–Love plate under quasi-static loading, PINN configurations with varying combinations of data, boundary conditions and governing partial differential equations are systematically compared. The results show that enforcing force and displacement boundary conditions, supplemented by sparse deflection sensors, can reconstruct plate deflections and bending moments with low bias, while explicit PDE terms further improve robustness to noise and modelling imperfections. For simply supported beams under moving loads, forward PINNs are developed using spatial–temporal inputs, and then extended to include axle load magnitude as an additional input. These models accurately reproduce deflection, moments and strains for loads both within and beyond the training range and demonstrate virtual sensing capabilities at locations that there are no physically sensors instrumented or it is not easy to measure the structural response. Inverse PINNs are then used to estimate flexural rigidity from synthetic and field strain data, achieving flexural-rigidity errors typically below 1% for noisy synthetic data and reducing the error from about 14% to less than 2% when real bridge measurements from a single sensor are incorporated. Finally, a multi-PINN framework for integrating PINNs based on domain decomposition is introduced in which a beam is partitioned into subdomains, each represented by its own network and stiffness parameter, with compatibility and equilibrium enforced at interfaces. Forward models predict full-field responses for beams discretised into multiple segments, while inverse models identify subdomain level stiffness reductions with relative errors generally below 3%, successfully localising both single- and multi-domain stiffness losses. Overall, the thesis demonstrates that PINNs provide a physically interpretable approach for bridge SHM, capable of fusing sparse monitoring data with governing physics to support virtual sensing and subdomain level system identification under moving loads.&lt;p&gt;&lt;/p&gt;","abstract_has_math":false,"creators":["Anmar Al-Adly (21049778)"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-06-08T00:00:00Z","date_published":"2026-06-08T00:00:00Z","updated_at":"2026-07-27T19:32:43Z","subjects":["SHM","PINNs","Bridges","Moving loads","Digital twin"],"languages":[],"rights":["All rights reserved","Open Access after 2027-06-08"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["10779/exe.32603244.v1"],"render_values":[{"text":"10779/exe.32603244.v1","href":null,"code":true}]}]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Anmar Al-Adly (21049778)"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026-06-08T00:00:00Z"]},{"key":"dc:relation","label":"Dc Relation","values":["https://figshare.com/articles/thesis/Physics-Informed_Neural_Networks_for_Structural_Health_Monitoring_of_Bridges/32603244"]},{"key":"dc:type","label":"Dc Type","values":["Text","Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["SHM","PINNs","Bridges","Moving loads","Digital twin"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["All rights reserved","Open Access after 2027-06-08"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["10779/exe.32603244.v1"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Bridges are critical components of a nation's transport infrastructure. Structural health monitoring (SHM) is increasingly employed, particularly on key bridges, to track their structural performance and for their effective structural management. Interpretation of sensing data from SHM however remains a challenge. Conventional data-driven models often require large volumes of measurements and lack physical interpretability, while purely physics-based models can be difficult to update with sparse, noisy monitoring data. This thesis investigates physics-informed neural network (PINN)-based frameworks for interpreting bridge monitoring data, with the overarching aim of enabling physically consistent prediction of structural response and identification of key stiffness parameters using limited measurements. As PINNs are a relatively new development with almost no prior application to SHM at the commencement of this study, this research's approach has been incremental, starting from simple scenarios, and gradually ramping up complexity, examining both forward and inverse problems. A general PINN methodology is first formulated for both forward and inverse problems, including the construction of physics-based loss functions, adaptive loss weighting, collocation points sampling strategies and error metrics for model evaluation. This framework is then applied to three classes of structural systems. For a Kirchhoff–Love plate under quasi-static loading, PINN configurations with varying combinations of data, boundary conditions and governing partial differential equations are systematically compared. The results show that enforcing force and displacement boundary conditions, supplemented by sparse deflection sensors, can reconstruct plate deflections and bending moments with low bias, while explicit PDE terms further improve robustness to noise and modelling imperfections. For simply supported beams under moving loads, forward PINNs are developed using spatial–temporal inputs, and then extended to include axle load magnitude as an additional input. These models accurately reproduce deflection, moments and strains for loads both within and beyond the training range and demonstrate virtual sensing capabilities at locations that there are no physically sensors instrumented or it is not easy to measure the structural response. Inverse PINNs are then used to estimate flexural rigidity from synthetic and field strain data, achieving flexural-rigidity errors typically below 1% for noisy synthetic data and reducing the error from about 14% to less than 2% when real bridge measurements from a single sensor are incorporated. Finally, a multi-PINN framework for integrating PINNs based on domain decomposition is introduced in which a beam is partitioned into subdomains, each represented by its own network and stiffness parameter, with compatibility and equilibrium enforced at interfaces. Forward models predict full-field responses for beams discretised into multiple segments, while inverse models identify subdomain level stiffness reductions with relative errors generally below 3%, successfully localising both single- and multi-domain stiffness losses. Overall, the thesis demonstrates that PINNs provide a physically interpretable approach for bridge SHM, capable of fusing sparse monitoring data with governing physics to support virtual sensing and subdomain level system identification under moving loads.<p></p>"]},{"key":"dc:title","label":"Title","values":["Physics-Informed Neural Networks for Structural Health Monitoring of Bridges"]}]}],"canonical_facts":{"dc:creator":["Anmar Al-Adly (21049778)"],"dc:date":["2026-06-08T00:00:00Z"],"dc:description":["Bridges are critical components of a nation's transport infrastructure. Structural health monitoring (SHM) is increasingly employed, particularly on key bridges, to track their structural performance and for their effective structural management. Interpretation of sensing data from SHM however remains a challenge. Conventional data-driven models often require large volumes of measurements and lack physical interpretability, while purely physics-based models can be difficult to update with sparse, noisy monitoring data. This thesis investigates physics-informed neural network (PINN)-based frameworks for interpreting bridge monitoring data, with the overarching aim of enabling physically consistent prediction of structural response and identification of key stiffness parameters using limited measurements. As PINNs are a relatively new development with almost no prior application to SHM at the commencement of this study, this research's approach has been incremental, starting from simple scenarios, and gradually ramping up complexity, examining both forward and inverse problems. A general PINN methodology is first formulated for both forward and inverse problems, including the construction of physics-based loss functions, adaptive loss weighting, collocation points sampling strategies and error metrics for model evaluation. This framework is then applied to three classes of structural systems. For a Kirchhoff–Love plate under quasi-static loading, PINN configurations with varying combinations of data, boundary conditions and governing partial differential equations are systematically compared. The results show that enforcing force and displacement boundary conditions, supplemented by sparse deflection sensors, can reconstruct plate deflections and bending moments with low bias, while explicit PDE terms further improve robustness to noise and modelling imperfections. For simply supported beams under moving loads, forward PINNs are developed using spatial–temporal inputs, and then extended to include axle load magnitude as an additional input. These models accurately reproduce deflection, moments and strains for loads both within and beyond the training range and demonstrate virtual sensing capabilities at locations that there are no physically sensors instrumented or it is not easy to measure the structural response. Inverse PINNs are then used to estimate flexural rigidity from synthetic and field strain data, achieving flexural-rigidity errors typically below 1% for noisy synthetic data and reducing the error from about 14% to less than 2% when real bridge measurements from a single sensor are incorporated. Finally, a multi-PINN framework for integrating PINNs based on domain decomposition is introduced in which a beam is partitioned into subdomains, each represented by its own network and stiffness parameter, with compatibility and equilibrium enforced at interfaces. Forward models predict full-field responses for beams discretised into multiple segments, while inverse models identify subdomain level stiffness reductions with relative errors generally below 3%, successfully localising both single- and multi-domain stiffness losses. Overall, the thesis demonstrates that PINNs provide a physically interpretable approach for bridge SHM, capable of fusing sparse monitoring data with governing physics to support virtual sensing and subdomain level system identification under moving loads.<p></p>"],"dc:identifier":["10779/exe.32603244.v1"],"dc:relation":["https://figshare.com/articles/thesis/Physics-Informed_Neural_Networks_for_Structural_Health_Monitoring_of_Bridges/32603244"],"dc:rights":["All rights reserved","Open Access after 2027-06-08"],"dc:subject":["SHM","PINNs","Bridges","Moving loads","Digital twin"],"dc:title":["Physics-Informed Neural Networks for Structural Health Monitoring of Bridges"],"dc:type":["Text","Thesis"]},"updated_at":"2026-07-27T19:32:43Z"}