Introduction
Structural Health Monitoring (SHM) is critical for ensuring the safety and longevity of infrastructure. Traditionally, SHM approaches have been divided into two distinct paradigms: statistical models that identify patterns in sensor data without physical context, and physics-based models (such as Finite Element Methods) that rely on idealized equations but often struggle with real-world noise and environmental variability. This paper proposes a "data-centric engineering" framework that bridges this gap by using multi-output Gaussian processes (GPs) to fuse sparse, noisy sensor data with the underlying physics of structural behavior.
Methodology: Fusing Physics with Gaussian Processes
The core of the proposed approach is a multi-output GP joint model. Unlike standard interpolation methods, this model treats the physical laws and the observed data as related outputs of the same underlying system.
The researchers utilize the Euler-Bernoulli beam equation, which describes the relationship between a beam's vertical deflection ($u$) and its curvature ($f$) through a linear differential operator. By embedding this operator directly into the covariance kernel of the GP, the model ensures that its predictions are mathematically consistent with structural mechanics.
A critical component of this framework is the balancing of simulation points ($N_u$). The model dynamically adjusts its reliance: if the physics model is overly simplified, it leans more on the empirical data; if the data is highly noisy, it prioritizes the physical constraints. Parameters such as flexural rigidity ($EI$) and ballast stiffness ($k$) are estimated using Maximum A Posteriori (MAP) estimation.
Experimental Validation
The methodology was validated using both simulated data and experimental results from a full-scale laboratory test of a prestressed concrete railway sleeper. The sleeper was instrumented with Fibre Bragg Grating (FBG) sensors and placed on a compacted railway ballast. It was subjected to increasing loads to simulate operational stress and eventually induce structural failure. The sensors provided microstrain data, which is inherently noisy and sparse.
Key Insights and Results
The GP framework demonstrated several key advantages over traditional methods:
- Inferring Hidden Parameters: The model successfully estimated parameters that are difficult to measure directly, such as the stiffness of the ballast ($k \approx 219.41$) and the sleeper's flexural rigidity ($EI \approx 9.07 \times 10^{11}$).
- Early Damage Detection: By monitoring the log marginal likelihood, the system can detect when the incoming data no longer aligns with the physics-based model. In the laboratory test, the model identified a structural change-point between 1100 and 1120 seconds, precisely matching the moment physical cracking was observed.
- Sensitivity: ROC analysis indicated an 80% probability of detecting a 10% decrease in structural parameters—changes that would be invisible to visual inspection or simple data thresholds.
Conclusion
This research provides a principled way to integrate "structural logic" into probabilistic data models. While currently limited to 1D Euler-Bernoulli models, the framework lays the groundwork for more complex "Digital Twins" using Finite Element Models. By identifying exactly where new sensors would provide the most value, this approach also offers a path toward optimized, active learning-based structural monitoring.
Summary updated