Real-time structural integrity assessment of bridge infrastructure

Sanduni Jayasinghe, Thathsara Nanayakkara, Azadeh Alavi, Mojtaba Mahmoodian, Amir Sidiq, Zhongchao Sun, Farham Shahrivar, Sujeeva SetungeOpenalex Unknown 2024 · DOI · OpenAlex

Abstract

Ageing infrastructure systems are easily prone to failure demanding frequent monitoring despite costs and risks. Digital twin offers an intelligent approach to overcome such limitations of traditional bridge monitoring. Current study focuses on implementing remotely connected real-time virtual models of bridges to visualize any induced stresses/strains of the actual bridge via the digital twin concept. It was first proposed to instrument a bridge with sensors and to follow an inverse approach to simulate the overall behavior using sensor measurements. A machine learning algorithm based finite element model has been suggested for this and three different algorithms were tested. All three machine learning algorithms accurately represented the behavior of the actual bridge in real-time. Overall, it was concluded that a combination of both physics-based and data-driven model development techniques is highly effective for real-time analyses and for the digital twining process. The proposed method can be extended to a large-scale bridge easily.

Introduction

Bridges are critical components of modern infrastructure, yet they face significant challenges due to aging and environmental degradation. In Australia alone, road network maintenance costs reach nearly $19 billion annually. Traditional structural health monitoring (SHM) systems are often reactive and computationally expensive, making real-time assessment difficult for large-scale assets.

While "Digital Twins"—virtual replicas of physical assets—have emerged as a promising solution, many existing implementations are limited to "geometric" replicas that lack the ability to calculate internal stresses or strains dynamically. The Finite Element Method (FEM) is the gold standard for such calculations but is too computationally intensive for true real-time applications. This research proposes a "structural" digital twin framework that leverages machine learning (ML) surrogates to provide instantaneous structural health assessments.

Methodology

The researchers developed a twofold methodology: training an ML surrogate model using synthetic data and deploying it for real-time visualization.

The Physical and Inverse Model

A laboratory-scaled stainless steel truss bridge was used as the physical testbed, instrumented with ten strain gauges (labeled SG_1 to SG_10). Unlike traditional forward modeling, which predicts reactions to known loads, this study employs an Inverse Method. Here, real-time sensor data (strain) serves as the input, and the model predicts the internal state, including nodal displacement and stress across the entire structure.

Schematic of the laboratory-scaled truss bridge showing dimensions and sensor locations (SG_1 to SG_10).

ML Surrogate Workflow

To overcome the latency of traditional FEM, three ML algorithms were tested as surrogates: Random Forest (RF), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). The workflow involves generating a large synthetic dataset from a validated FE model to train the ML algorithms. Once trained, these models can process live sensor readings in milliseconds, bypassing the need for heavy FE computations during monitoring.

Methodological flowchart showing (a) the ML training process using FE data and (b) the real-time monitoring and visualization pipeline.

Data and Training

The study utilized a synthetic dataset of 4,000 records generated via automated Python scripting of the FE model. The dataset used strain values as inputs and nodal displacements as outputs, with an 80/20 split for training and validation. By using synthetic data for training, the researchers ensured the ML model understood the underlying physics of the truss before it ever processed a real sensor reading.

Key Results and Discussion

The integration of ML surrogates achieved a significant breakthrough in processing speed without sacrificing accuracy.

  • Accuracy: All three algorithms demonstrated high precision. Random Forest emerged as the top performer with a Root Mean Square Error (RMSE) of 0.000341, followed by XGBoost (0.000460) and MLP (0.000473).
  • Latency: The system achieved a processing time of approximately 20 milliseconds per update. This is 50 times faster than the 1-second refresh rate of the sensors, enabling true real-time monitoring.
  • Visualization: The output was integrated into a Python-based tool that provides a color-coded visualization of the bridge members (e.g., red for compression, blue for tension), allowing operators to see exactly which components are under the most stress at any given second.

Conclusion

This research demonstrates that combining physics-based FE modeling with data-driven ML techniques is highly effective for real-time structural analysis. By moving from geometric to structural digital twins, engineers can achieve a deeper understanding of how infrastructure "feels" loads in real-time. While the current study focused on a laboratory scale, the framework is designed for scalability to full-sized operational bridges, potentially revolutionizing how we maintain and monitor critical infrastructure.

Summary updated

Abstract

Ageing infrastructure systems are easily prone to failure demanding frequent monitoring despite costs and risks. Digital twin offers an intelligent approach to overcome such limitations of traditional bridge monitoring. Current study focuses on implementing remotely connected real-time virtual models of bridges to visualize any induced stresses/strains of the actual bridge via the digital twin concept. It was first proposed to instrument a bridge with sensors and to follow an inverse approach to simulate the overall behavior using sensor measurements. A machine learning algorithm based finite element model has been suggested for this and three different algorithms were tested. All three machine learning algorithms accurately represented the behavior of the actual bridge in real-time. Overall, it was concluded that a combination of both physics-based and data-driven model development techniques is highly effective for real-time analyses and for the digital twining process. The proposed method can be extended to a large-scale bridge easily.
BibTeX
@inproceedings{jayasinghe2024real,
  author = {Sanduni Jayasinghe and Thathsara Nanayakkara and Azadeh Alavi and Mojtaba Mahmoodian and Amir Sidiq and Zhongchao Sun and Farham Shahrivar and Sujeeva Setunge},
  title = {{Real-time structural integrity assessment of bridge infrastructure}},
  booktitle = {Openalex Unknown},
  year = {2024},
  doi = {10.1201/9781003483755-288},
}
Figure