⚡ electrical engineering

Modal Decomposition and Identification for a Population of Structures Using Physics-Informed Graph Neural Networks and Transformers

This study proposes a novel, unsupervised deep learning framework that integrates graph neural networks and transformers with a physics-informed loss function to accurately decompose structural dynamic responses and identify modal properties across a population of structures without requiring labeled data.

Original authors: Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xudong Jian, Kiran Bacsa, Gregory Duthé, Eleni Chatzi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to diagnose a patient's health. In the world of civil engineering, the "patients" are bridges, skyscrapers, and wind turbines. To keep them safe, engineers need to listen to how they "vibrate" or "sing" when the wind blows or cars drive over them. This singing is called modal identification.

However, listening to a giant bridge is tricky. You can't put a microphone on every single bolt (that's too expensive), and the "song" is often a messy mix of many different notes playing at once.

This paper introduces a new, super-smart "AI Doctor" that can listen to a few microphones, figure out the entire song, and tell you exactly which notes are healthy and which might be broken. Here is how it works, broken down into simple concepts:

1. The Problem: The "Messy Choir"

Imagine a massive choir (the bridge) singing a complex song.

The Issue: You only have microphones on a few singers (sparse sensors).
The Noise: The wind and traffic are shouting over them.
The Mix: The song is a jumbled mess of different voices (modes) singing together. Traditional methods try to untangle this mess, but they often get confused, especially if they haven't seen that specific choir before.

2. The Solution: A Team of AI Specialists

The authors built a new AI system that acts like a team of three specialists working together:

Specialist A: The "Map Reader" (Graph Neural Networks)

The Analogy: Imagine you have a map of a city, but you only know the traffic flow on a few main streets. A normal computer might get lost. This specialist, however, understands the shape of the city. It knows that if a street connects to a busy avenue, the traffic there is likely related.
What it does: It looks at the physical shape of the bridge (the graph) and uses the few sensors you have to "fill in the blanks." It predicts what the vibration looks like on the parts of the bridge where you have no sensors. It turns a partial puzzle into a complete picture.

Specialist B: The "Conductor" (Transformers)

The Analogy: Imagine the choir is singing a chaotic jam session. The Conductor steps in and says, "Stop! Let's separate the voices."
What it does: The bridge's vibration is a mix of many different "notes" (frequencies) happening at once. This specialist uses a powerful attention mechanism to listen to the messy signal and split it apart into individual, pure notes (Single-Degree-of-Freedom modes). It isolates the "bass" from the "treble" so you can analyze them separately.

Specialist C: The "Physics Detective" (Physics-Informed Loss)

The Analogy: Usually, to train a student, you give them a test with the answers (labeled data). But in real life, you don't have the answers! You don't know the "true" vibration of a bridge until you've already measured it.
What it does: Instead of needing an answer key, this specialist teaches the AI using rules of physics. It tells the AI: "Hey, different notes in a choir shouldn't interfere with each other; they should be independent." If the AI tries to mix the notes, the Detective says, "No, that breaks the laws of physics!" This allows the AI to learn perfectly without ever seeing the "correct" answer beforehand.

3. The "Population" Superpower

Most AI models are like a student who studies one specific bridge and fails when they see a slightly different one.

The Innovation: This AI is trained on a population of bridges. It learns the general "rules of bridge-singing" rather than memorizing one specific bridge.
The Result: Once trained, it can walk up to a new bridge it has never seen before, listen to a few sensors, and immediately start diagnosing it. It's like a doctor who learns general anatomy and can treat any patient, not just the ones they practiced on.

4. Why is this a Big Deal?

No Labels Needed: It doesn't need expensive, pre-measured data to learn. It learns by understanding physics.
Fills in the Gaps: It can reconstruct the vibration of the entire bridge even if you only have sensors on 18% of it.
Better than the Old Ways: The paper tested this against traditional methods (like EFDD and SSI). The AI was better at finding the "high notes" (complex vibrations) and was more accurate at spotting the shape of the vibrations, especially when the data was messy or incomplete.
Real-World Ready: They tested it on a real, scaled-down cable-stayed bridge in a lab with different types of "traffic" (pulling weights and moving toy cars), and it worked great.

The Catch (Limitations)

The AI is currently a bit slow to "study" (train) because it's doing a lot of math. It's like a brilliant student who takes a long time to study for an exam but then solves the test perfectly. Once trained, it's fast, but the initial training takes time. Also, while it's great at generalizing to new bridges, it still needs a little bit of re-learning if the new bridge is very different or if the "traffic" (excitation) is extremely chaotic.

Summary

This paper presents a new "AI Doctor" for bridges. It uses a Map Reader to guess missing data, a Conductor to separate mixed-up vibrations, and a Physics Detective to ensure the results make sense without needing a cheat sheet. It's a smarter, more flexible way to keep our infrastructure safe and healthy.

1. Problem Statement

The paper addresses critical challenges in Structural Health Monitoring (SHM) and Operational Modal Analysis (OMA), specifically within the context of Population-Based SHM (PBSHM). The core problems identified are:

Data Scarcity and Sparsity: In real-world scenarios, sensors are often limited, providing measurements for only a fraction of structural nodes ( $M \ll N$ ), while full-field mode shapes are required.
Lack of Labeled Data: Traditional supervised deep learning requires labeled ground-truth modal parameters (frequencies, damping, mode shapes), which are rarely available in operational settings.
Generalization: Existing methods struggle to generalize across a population of structures with varying geometries, topologies, and material properties.
Mode Separation: Conventional methods (e.g., FDD, SSI) often suffer from mode mixing, end effects, and difficulty separating complex multi-degree-of-freedom (MDOF) responses into single-degree-of-freedom (SDOF) modes without physical constraints.

The goal is to develop a framework that can decompose sparse, unlabeled dynamic measurements into SDOF modal responses and identify modal properties (frequencies, damping, mode shapes) for a population of diverse structures without requiring prior labeled data.

2. Methodology

The proposed framework integrates Graph Neural Networks (GNNs), Transformers, and Physics-Informed Learning into a unified unsupervised pipeline.

A. Feature Propagation (FP) for Full-Field Reconstruction

To address sparse sensor data, the authors employ a Feature Propagation algorithm.

Input: Sparse acceleration data from a subset of nodes and the structural topology (adjacency matrix).
Process: The algorithm iteratively propagates known measurements to unknown nodes based on graph connectivity.
Output: Reconstructed full-field dynamic measurements ( $\tilde{\mathbf{X}}$ ) for all $N$ nodes, enabling the model to learn from complete spatial data despite sparse inputs.

B. Deep Learning Architecture

The core model consists of three distinct blocks designed to handle graph-structured and temporal data:

GraphSAGE (GNN): Processes the reconstructed full-field data. It treats the structure as a graph where nodes are joints and edges are elements. GraphSAGE aggregates neighbor information to extract spatial features, handling varying node counts and topologies inherent in a population of structures.
Set Transformer: Converts the spatial features from GraphSAGE into time-domain SDOF modal responses ( $\hat{\mathbf{Q}}$ ). Unlike standard Transformers, the Set Transformer is permutation-invariant, making it suitable for graph-level tasks where the number of nodes varies.
Multi-Layer Perceptron (MLP): Projects the hidden features back to the node level to predict mode shapes ( $\hat{\mathbf{\Phi}}$ ) for all structural nodes.

C. Physics-Informed Loss Function

Since no labeled data exists, the model is trained using an unsupervised, physics-informed loss function ( $\mathcal{L}$ ) comprising three terms:

Reconstruction Loss: Minimizes the Mean Squared Error (MSE) between the reconstructed full-field data and the superposition of predicted mode shapes and responses ( $\hat{\mathbf{\Phi}} \cdot \hat{\mathbf{Q}} \approx \tilde{\mathbf{X}}$ ).
Time-Domain Independence: Enforces that the time-domain responses of different modes are uncorrelated (orthogonal). The correlation matrix of $\hat{\mathbf{Q}}$ is penalized if it deviates from an identity matrix.
Frequency-Domain Independence: Enforces that the frequency spectra (via FFT) of different modes are distinct. The correlation matrix of the amplitude spectra is penalized to ensure modes do not overlap in frequency.

Once trained, natural frequencies and damping ratios are extracted from the predicted SDOF responses using standard techniques like Power Spectral Density (PSD) and Random Decrement Technique (RDT).

3. Key Contributions

Novel Architecture: Integration of GraphSAGE and Set Transformers to simultaneously handle structural topology (spatial) and dynamic responses (temporal) for a population of structures.
Label-Free Training: A physics-informed loss function that eliminates the need for labeled ground-truth modal parameters, relying solely on the physical principles of modal decomposition and mode independence.
Sparse-to-Full Reconstruction: A feature propagation mechanism that leverages structural topology to reconstruct full-field mode shapes from sparse sensor data, overcoming a major limitation in conventional SHM.
Population Generalization: A framework capable of learning from a training set of structures and applying modal decomposition to unseen structures within the same population, even under varying loading conditions.

4. Experimental Results

The method was validated through numerical simulations on a heterogeneous population of trusses and laboratory experiments on a scaled cable-stayed bridge.

Numerical Experiments (Truss Population)

Performance: The model successfully decomposed MDOF responses into SDOF modes and identified natural frequencies and damping ratios.
Comparison: Compared against Enhanced Frequency Domain Decomposition (EFDD) and Stochastic Subspace Identification (SSI):
- Accuracy: Comparable accuracy for low-order modes and frequencies/damping.
- High-Order Modes: Significantly superior performance in identifying high-order mode shapes (higher MAC values) because the model utilizes structural topology, whereas EFDD/SSI rely on interpolation.
- Efficiency: While training is computationally expensive, inference is fast. The model provides a unique capability to separate individual mode responses, which EFDD/SSI cannot do directly.
Ablation Study: Removing the GNN, Transformer, or physics-informed constraints resulted in significant performance degradation, confirming the necessity of each component.

Laboratory Experiments (Cable-Stayed Bridge)

Setup: Tests included "pull-and-release" (free decay) and "vehicle load" (forced vibration) scenarios.
Results:
- The model achieved high accuracy (MAC > 0.95) for modes 1 and 2 across most tests.
- Performance was robust for free decay vibrations but showed reduced generalization for complex, non-stationary vehicle loads (Test 10), highlighting current limitations in handling highly complex excitations without retraining.
- Outperformed SSI in damping ratio estimation stability and matched EFDD in frequency estimation.

5. Significance and Limitations

Significance:

Practical Applicability: By removing the need for labeled data and handling sparse sensors, the method is highly applicable to real-world OMA where ground truth is unknown.
Population-Based Approach: It advances the field of PBSHM by demonstrating that deep learning models can generalize across structural populations, not just single structures.
Mode Separation: Unlike traditional methods that provide global modal parameters, this approach explicitly decomposes the signal into individual SDOF time-domain responses, offering deeper insights for vibration control and damage detection.

Limitations:

Generalization to New Structures: While the model generalizes well to new measurements of the same structure or similar topologies, its ability to generalize to entirely new structural geometries without retraining is limited.
Complex Excitations: Performance degrades under complex, non-stationary excitations (e.g., moving vehicle loads) compared to simple free-decay vibrations.
Computation Time: The training phase is computationally intensive compared to classical methods, though inference is fast.

Conclusion:
This paper presents a robust, physics-informed deep learning framework that bridges the gap between data-driven methods and physical structural dynamics. It offers a scalable solution for modal identification in populations of structures, effectively handling sparse data and unlabeled scenarios, marking a significant step forward in automated Structural Health Monitoring.