AI Reveals Hidden Weaknesses in Bridge Monitoring

The sudden collapse of the Carola Bridge in Dresden in 2024 highlighted a critical gap in how we monitor aging infrastructure. While structural health monitoring (SHM) systems collect data from bridges worldwide, they often fail to quantify the reliability of their inferences, leaving decision-makers uncertain about actual structural capacity. This problem is particularly acute for prestressed concrete bridges, where hidden issues like tendon corrosion and stress-corrosion cracking can degrade flexural rigidity—the key property dictating how loads translate to deformation—long before visible damage appears. A new study from Technische Universität Dresden addresses this by developing a Bayesian inverse framework that not only estimates these hidden parameters but rigorously quantifies where and why uncertainty persists, offering a more transparent approach to bridge safety.

Researchers discovered that traditional deterministic s for inferring flexural rigidity from rotation measurements, while useful, obscure critical limitations in identifiability. Using Fisher information as a diagnostic metric, they found that sensor informativeness varies dramatically along a bridge span. For example, in a simply supported beam, sensors placed at one-quarter of the span length provide maximal information about flexural rigidity near one-third of the span, but very little near the supports. This spatial heterogeneity means that certain regions, especially near zero-moment zones where bending moments vanish, are inherently weakly identifiable from rotation data alone. The Bayesian framework makes these limits explicit by producing credible intervals that widen in low-information areas, revealing where estimates are least reliable.

Ology applies a Bayesian inverse problem approach to recover distributed flexural rigidity from rotation influence lines measured during controlled vehicle passages. The forward model uses Euler-Bernoulli beam theory, discretizing the span into elements and relating rotations to compliance through a linear system. A Gaussian likelihood incorporates measurement noise, while Gaussian Markov random field priors enforce smoothness and positivity on the flexural rigidity field. This formulation unifies classical Tikhonov regularization with uncertainty quantification, as the maximum a posteriori estimate corresponds to the Tikhonov solution, and the posterior covariance supplies uncertainty bounds. The researchers computed per-sensor Fisher information to assess how specific sensor layouts constrain recoverable spatial features, with the information decomposing additively across sensors.

From synthetic studies and application to the full-scale openLAB research bridge demonstrate the framework's capabilities. In synthetic tests, as measurement noise decreased, posterior credible intervals contracted, allowing sharper recovery of damaged zones, but uncertainty remained elevated near supports regardless of noise level. Figure 4 shows a posterior flexural rigidity profile with spatially heterogeneous uncertainty, where bands are widest near supports and narrower in mid-span regions. Figure 7 illustrates the bias-variance trade-off, where total error at mid-span decreases with finer discretization until over-parameterization amplifies variance, with optimal performance around 40-60 elements for given sensor counts. On the openLAB bridge, using data from two tiltmeter stations, identified span 1 as less stiff than span 2, consistent with material differences, but credible intervals widened significantly near the right-side support, reflecting reduced sensor informativeness in that zero-moment region.

For real-world bridge management are substantial. By quantifying uncertainty, this approach provides asset managers with transparent evidence for load-rating updates and digital-twin calibration, distinguishing between well-informed estimates and guesses. It also offers practical guidance for sensor placement: for example, sensors should be positioned to cover both sides of interior supports in continuous spans, as shown in Figure 3, where upstream sensors retain diminishing sensitivity beyond supports. However, the study acknowledges limitations, including reliance on small-deflection Euler-Bernoulli kinematics, quasi-static loading, and log-normal priors that assume smoothness. It omitted dynamic effects, temperature drifts, torsional coupling, and model-discrepancy terms, which may underestimate epistemic uncertainty. Future work aims to fuse rotations with complementary data like displacements or strains and implement online inference for sequential updating, potentially enhancing resilience assessments for aging infrastructure worldwide.