How do you spot a meaningful change in a bridge before it becomes a maintenance problem?
- Context
- COWI · Clifton Suspension Bridge
- My contribution
- Anomaly-detection modelling, dashboard, automated alerts and reports
- System
- Structural health monitoring
- Public record
- 2026 engineering case study
The movement that matters
The bridge deck hangs from chains that pass over roller-mounted saddles at the top of each tower. Those saddles need to move as traffic, wind and temperature change the forces across the structure.
If their movement gradually deteriorates, loads can be transferred into the tower masonry and foundations in ways the system was not designed to carry. The monitoring problem is therefore not simply “did a sensor move?” It is whether the observed movement still makes sense for the conditions.
Conceptual reconstruction
Expected movement.
Observed movement.
The gap between them.
- Saddledisplacement and plane rotation
- Chainsstrain and temperature
- Environmentambient temperature
- Decisionmaintenance warning
My part of the system
I modelled the anomaly-detection layer and built the interface around it: a dashboard for seeing the bridge’s behaviour, automated alerts when patterns needed attention, and reports that made the monitoring history reviewable.
The important shift was from displaying sensor values to providing context. A useful system had to distinguish normal temperature-driven movement from a deviation worth investigating, then communicate that difference without asking an engineer to continuously watch raw streams.
From a signal
to a decision.
The product layer turned continuous monitoring into a repeatable maintenance workflow.
- 01
Collect
Bring displacement, rotation, strain and temperature streams into one pipeline.
- 02
Model
Estimate how the saddles should move under the measured conditions.
- 03
Compare
Look for departures from expected behaviour rather than isolated high values.
- 04
Act
Surface the pattern through the dashboard, alerts and periodic reporting.
What the public record confirms
The published engineering case study describes a fully automated process that collected and processed monitoring data, predicted saddle displacement and rotation, and compared actual movement with those predictions.
At publication, linear regression was being used to identify gross physical changes such as roller fracture or seizure. The article also describes work on more sensitive models intended to detect gradual deterioration, including wear-related changes in articulation, early enough for targeted intervention.
Engineering software is decision support
The technically interesting part was the model. The professionally important part was the whole path from sensor to maintenance decision.
This project taught me to treat the dashboard, alerts and reports as part of the monitoring system itself. Detection is only valuable when the right person can understand what changed, see the evidence, and decide what should happen next.