Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Mechanical Engineering | en_US |
| dc.contributor.advisor | Cheng, Li (ME) | en_US |
| dc.creator | Song, Yang | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/14527 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Physics-informed neural network for robust time-of-flight extraction in complex wavefields : applications in guided-wave-based structural health monitoring | en_US |
| dcterms.abstract | In wave-based sensing technologies, accurately extracting Time-of-Flight (ToF) is a universally challenging scientific problem, especially in fields like acoustics, electromagnetics, and elastodynamics. In complex environments, multipath interference and scattering from boundaries, obstacles, and various structural features cause severe signal superposition. This makes it difficult for traditional signal processing and purely data-driven methods to precisely isolate the target wave packet, significantly reducing the sensing system's accuracy and reliability. This thesis focuses on guided-wave technology in Structural Health Monitoring (SHM) as a key application and systematically explores a series of innovative Physics-Informed Machine Learning (PIML) frameworks to address the challenge of robust ToF extraction in complex wavefields. | en_US |
| dcterms.abstract | The research in this thesis progresses systematically. First, to address the primary source of interference in SHM, boundary reflections, a Multi-Scale Spatio-Temporal (MSST) fusion network is introduced. Using a physics-guided architectural design, this network combines the temporal information of signals with the spatial geometry of sensors, effectively learning to remove the influence of boundary reflections on damage-scattered signals. Interpretability analyses, including deconvolution-based visualization, confirm the network's ability to effectively separate the target signal, providing a practical solution to expand the effective inspection area of structures. | en_US |
| dcterms.abstract | To further improve model interpretability and explicitly simulate the physical processes of wave phenomena, this dissertation introduces the Physics-informed Ray-tracing augmented Graph Neural Network (PhysRayGNN). This framework attempts to use a Graph Neural Network (GNN) to model wave propagation, attenuation, and superposition in SHM. By integrating a physics-based ray-tracing algorithm to build a high-fidelity wave-path graph and employing an attention-based GNN to simulate wave dynamics on this graph, the model's internal decision-making process aligns closely with physical laws. The learning process is also regularized by physics-informed constraints derived from the wave equation and wavelet features, ensuring the physical consistency of the predictions. Experimental results show that the framework can accurately disentangle highly superimposed signals, achieving centimeter-level damage localization precision. | en_US |
| dcterms.abstract | Finally, to tackle the challenge of real-world engineering structures with many complex scatterers, such as stiffeners and bolts, a more comprehensive and resilient multimodal framework, WaveSenseNet, is developed. This framework features several key innovations: (1) a new parametric activation function designed to address the inherent "low-frequency bias" of typical neural networks, effectively capturing high-frequency wave packet features essential for ToF; (2) a Gaussian positional encoding map that represents complex environmental geometry and scatterer physics as spatial features, enabling an in-depth multimodal fusion of time-domain signals and spatial data; and (3) the use of Monte Carlo Dropout to provide uncertainty estimates for ToF predictions, giving the final localization outcomes probabilistic confidence. In complex experimental settings with multiple scatterers, WaveSenseNet outperforms existing top models, further confirming the benefits of a deep fusion of physics and data. | en_US |
| dcterms.abstract | In summary, this thesis offers a systematic, interpretable, and high-precision approach to ToF extraction in complex wavefields through a series of increasingly advanced PIML frameworks. The results not only significantly improve the reliability and intelligence of guided-wave SHM technology but also introduce a new paradigm for integrating prior physical knowledge into future machine learning models, providing substantial theoretical and practical value for the broader wave-based sensing field. | en_US |
| dcterms.extent | xvi, 267 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2026 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Elastic wave propagation | en_US |
| dcterms.LCSH | Structural health monitoring | en_US |
| dcterms.LCSH | Neural networks (Computer science) | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
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/200/14527