Structural health monitoring · continuous damage detection
A bridge under traffic repeats itself. The same responses, season after season, train after train. Model that normal condition precisely enough and you no longer need to know in advance what a problem looks like — anything that departs from normal announces itself, on the day it appears rather than at the next inspection. That is novelty detection.
Above: one point per vehicle passage, plotted in two damage-sensitive features. The amber region is the healthy condition, learned from a year of traffic. A single point outside it is noise. A cluster of them, forming in the same place over many passages, is the structure telling you something.
Supervised learning wants labelled failures. For a rail bridge that means data recorded from that structure, with that damage, under that traffic — and the only way to get it is for the damage to already have happened. The training set you need is the outcome you are paid to prevent.
Threshold alarms fail for a quieter reason. A single channel crossing a limit is a late signal; by then the change is large. The early signal is not in any one channel, it's in the relationship between them — how strain at one point moves with strain at another, how the modes shift together as the train passes. Damage breaks the relationship long before it breaks a limit.
Novelty detection works from the one dataset you always have: the structure behaving normally. Build the statistical description of that, and anything that doesn't fit it is, by definition, new. You never have to specify what you're looking for.
The method
Raw acceleration and strain are not the input. From each measurement window we compute quantities that respond to stiffness and load path — natural frequencies and mode shapes, transmissibility between sensor pairs, autoregressive model coefficients, rainflow-counted stress cycles. Good features move when the structure changes and stay still when nothing has.
Feed in a baseline period long enough to contain the structure's whole ordinary life — winter and summer, empty track and full freight, wind, sun on one flange. The result is a description of the region those feature vectors occupy: a covariance, a density, a decision boundary. Not a threshold per channel — one boundary in all of them at once.
Temperature moves modal frequencies by several percent across a year. Real damage often moves them by less. Before testing anything, the environmental and operational component is projected out — so the residual reflects the structure, not the season. This is the step that decides whether the system is useful or noise.
Every new feature vector gets a discordancy score — its distance from the healthy region, in the metric the baseline defines. The threshold comes from the baseline's own distribution, so the false-alarm rate is a number you choose rather than a number you discover. Scores are tracked on a control chart: isolated exceedances are noise, a sustained run is a structure telling you something.
The hard part
Most novelty detection deployed on structures fails here, and it fails in the expensive direction: alarms every autumn, silence when it matters.
What it surfaces
Fatigue
Local stiffness loss shifts the strain relationship between neighbouring gauges well before any single reading looks unusual.
Bearings & joints
A seizing bearing changes how the deck responds to temperature and to passing load. The signature is in the response, not in a displacement limit.
Foundations
Slow support-condition change alters mode shapes without moving any frequency far. Only a multivariate baseline sees it.
Events
Compare the structure to its own recorded self from the week before. The question "did anything change?" gets an answer with a number on it.
What makes it possible
Outlier analysis for damage detection was proposed in the 1990s. It became practical when continuous, synchronised, high-rate measurement from a live structure stopped being a research exercise. That is the part IoT Bridge builds.
| Capability | Specification | Why the method needs it |
|---|---|---|
| Sampling rate | up to 400 Hz | Resolves the modal content a passing train excites |
| Sensing | Strain · acceleration · inclination | Multivariate features need simultaneous, co-located channels |
| Edge gateway | Wired & wireless, onboard compute | Features extracted at the structure; bandwidth stays finite |
| Baseline | Continuous, rolling | A healthy condition learned across full seasonal range |
| Analytics | Damage detection · rainflow · spectra | Condition and remaining-life assessment on the same stream |
| Digital twin | Hosted per structure | Ties a discordant reading to a location on the bridge |
| Export | Open standard formats | Findings land in the asset management system you already run |
Track record
The algorithms are developed with structural engineering researchers, not adapted from a general-purpose anomaly product. Instrumentation, acquisition, analytics and reporting come from one place, because a feature is only as trustworthy as the channel underneath it.
Where this comes from
Talk to us
If it already carries sensors, we can tell you whether the data supports a baseline. If it doesn't, we instrument it. Either way the first conversation is with an engineer who has done this on a bridge under traffic.