Modal Decomposition and Identification for a Population of Structures Using Physics-Informed Graph Neural Networks and Transformers
Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi · 2025
Read Modal Decomposition and Identification for a Population of Structures Using Physics-Informed Graph Neural Networks and Transformers by Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-04018.txt.
Introduction
Jian and co-authors teach a neural network to read the vibrations of whole families of structures — bridges, frames, towers — and split the motion into its natural modes. Knowing a structure's modes is basic to checking its health, but standard methods handle one structure at a time. Their framework blends graph neural networks, which grasp each structure's layout, with transformers, which unmix the combined vibration signal into single-mode responses. A physics-based penalty trains the system without labeled data.
In simulations and lab experiments, the model recovered natural frequencies, damping ratios, and mode shapes accurately from sparse sensors. It held up when loads changed and when the structure itself varied, and it beat established methods in head-to-head comparisons. For engineers monitoring whole fleets of assets, the paper points to one model that can watch them all.
Catalog record
| CREATOR | Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-04018
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Modal Decomposition and Identification for a Population of Structures Using Physics-Informed Graph Neural Networks and Transformers Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi Abstract Modal identification is crucial for structural health monitoring and structural control, providing critical insights into structural dynamics and performance. This study presents a novel deep learning framework that integrates graph neural networks (GNNs), transformers, and a physics-informed loss function to achieve modal decomposition and identification across a population of structures. The transformer module decomposes multi-degrees-of-freedom (MDOF) structural dynamic measurements into single-degree-of-freedom (SDOF) modal responses, facilitating the identification of natural frequencies and damping ratios. Concurrently, the GNN captures the structural configurations and identifies mode shapes corresponding to the decomposed SDOF modal responses. The proposed model is trained in a purely physics-informed and unsupervised manner, leveraging modal decomposition theory and the independence of structural modes to guide learning without the need for labeled data. Validation through numerical simulations and laboratory experiments demonstrates its effectiveness in accurately decomposing dynamic responses and identifying modal properties from sparse structural dynamic measurements, regardless of variations in external loads or structural configurations. Comparative analyses against established modal identification techniques and model variations further underscore its superior performance, positioning it as a favorable approach for population-based structural health monitoring. [orcid=0000-0001-9973-9662] [orcid=0000-0002-0834-185X] [orcid=0000-0002-0895-6766] [orcid=0000-0002-6870-240X] † † highlights: Population-based modal identification framework . The proposed approach is tailored for population-based structural health monitoring (PBSHM), enabled by Graph Neural Networks (GNNs) that effectively model structural topologies with varying node configurations. This facilitates scalable modal identification across multiple structures under diverse loading conditions. Label-free modal decomposition guided by physics . By integrating Set Transformers and a physics-informed loss function grounded in modal decomposition theory, the model performs accurate decomposition of dynamic responses into single-degree-of-freedom (SDOF) modes without requiring labeled data, significantly enhancing practical applicability in real-world operational modal analysis (OMA). Full-field identification from sparse measurements . A feature propagation algorithm leverages structural topology to reconstruct full-field mode shapes and dynamic responses from sparse sensor data, addressing a key limitation in conventional SHM where measurements are often incomplete. Superior performance under realistic conditions . Validated through both numerical simulations and laboratory experiments, the proposed model outperforms traditional modal identification methods (e.g., EFDD, SSI) and standard deep learning models (e.g., MLPs, LSTMs) in identifying natural frequencies, damping ratios, and high-order mode shapes, particularly under non-stationary and data-sparse conditions. keywords population-based SHM ,modal decomposition ,modal identification ,physics-informed neural networks ,graph neural networks ,transformers , † † credit: Conceptualization, Methodology, Software, Data curation, Formal analysis, Data visualization, Writing - original draft, Writing - Review & Editing † † credit: Methodology, Software, Writing - Review & Editing † † credit: Methodology, Software, Writing - Review & Editing † † credit: Supervision, Funding Acquisition, Methodology, Project administration, Writing - Review & Editing 1 Introduction 1.1 Motivation Structural health monitoring (SHM) and vibration control are essential for ensuring the safety, functionality, and resilience of civil structures such as buildings, bridges, and wind turbines. As civil structures are increasingly subjected to dynamic loads such as traffic-induced vibrations, earthquakes, wind, and waves, the need for SHM and structural vibration control has become more critical. SHM enables real-time assessment of structural integrity, allowing for early detection of damage and deterioration, which is crucial for preventing catastrophic failures [ 1 ] . In parallel, structural vibration control techniques, including passive, active, and hybrid approaches, have been developed to mitigate excessive vibrations, enhancing the resilience and performance of structures under unwanted dynamic loading conditions [ 2 ] . A fundamental aspect of both SHM and structural vibration control is the ability to accurately capture and interpret the dynamic behavior of structures [ 3 , 4 ] . Modal decomposition and identification play a pivotal role in this process, as they provide insight into the modal properties of a structure [ 5 ] , including natural frequencies, mode shapes, damping characteristics, and modal response. For SHM, changes in these modal properties could serve as indicators of structural damage, making modal identification a widely-accepted tool for damage detection and health assessment [ 6 , 7 , 8 ] . Moreover, structural vibration control strategies generally rely on precise modal information to optimize control algorithms, ensuring effective vibration mitigation and stability enhancement [ 9 , 10 , 11 ] . Since modern civil structures become more complex and exposed to increasing risks, the development of more advanced modal decomposition and identification methods remains a critical area of research. 1.2 Related Work There have been extensive studies in modal identification for structures, which can be generally classified into Experimental Modal Analysis (EMA) or Operational Modal Analysis (OMA) techniques. Early modal identification techniques fall under EMA, which was developed over half a century ago to identify structural modal properties by applying controlled excitations (such as impact hammers or shakers) and measuring the response [ 12 ] . While effective in laboratory settings, EMA is often impractical for real-world civil structures due to inevitable environmental noise and the high costs associated with generating artificial excitation for large-scale systems. To overcome these limitations, OMA techniques were developed since 1980s to identify modal properties under actual operational conditions [ 13 ] . By capturing structural vibration responses induced by ambient sources such as wind and traffic, OMA avoids the need for expensive and intrusive testing, making it particularly valuable for large civil infrastructure. In this sense, OMA is also called output-only modal identification, as it only uses output responses without measuring input forces. Moreover, OMA techniques are generally categorized into frequency-domain and time-domain approaches, with Frequency Domain Decomposition (FDD) [ 14 ] and Stochastic Subspace Identification (SSI) [ 15 ] being the most widely used techniques in their respective domains. However, classic OMA techniques still face challenges in mode separation, automation, and robustness against noise and environmental variability. To address these challenges, many advanced OMA techniques have been developed over the past two decades, with a notable class of methods integrating signal decomposition techniques, such as independent component analysis (ICA) [ 16 ] and variational model decomposition (VMD) [ 17 ] , for modal identification. The core idea behind these approaches is to decompose the dynamic measurements of multi-degree-of-freedom (MDOF) structures into the dynamic responses of single-degree-of-freedom (SDOF) vibration modes while isolating measurement noise [ 18 ] . This not only simplifies modal identification, as identifying modal properties of SDOF systems is more straightforward, but also provides the dynamic response of structural vibration modes as a valuable byproduct, enabling applications in vibration control and structural design optimization. Nonetheless, these techniques still demonstrate limitations such as inadequate mode separation (mode mixing), end effects, lack of physical interpretability, and dependence on decomposition parameters [ 19 ] . Modal identification, like other inverse problems, is essentially an optimization problem that seeks to minimize the discrepancy between identified modal properties and their true values. As a powerful tool to solve complex optimization problems, a few studies have recently introduced machine learning and deep learning approaches to enhance modal decomposition and identification. Based on the model training approach, these studies can be categorized into two types: 1) Unsupervised learning-based approaches : Liu et al. [ 20 ] were among the first to apply machine learning to output-only modal decomposition and identification. By