Statistical methods for instrumented infrastructure

Today we are experiencing a rapid development of technologies such as robotics, Internet-of-Things and 3D printing. Instrumentation of such technologies and their related systems is already a part of this development and will revolutionise how infrastructure is delivered, maintained and controlled. For instance, engineers are starting to instrument railway bridges with fibre optic sensors, in order to understand more about bridges structural health. Through these instrumented structures, engineers are now encountering vast amounts of data and wish to understand the stochastic nature of the bridge. Further, they want, not only to understand a single instrumented bridge, but a network of bridges, to manage their operation, resilience and reliability.


Picture of instrumented bridge with a train passing over.

In this project, we will develop new statistical methods to improve productivity of networks through better understanding of the instrumented infrastructure. We will be working alongside the engineers in the Cambridge Centre for Smart Infrastructure and Construction that are monitoring railway bridges. This work may lead to a more accurate assessment of the capacity of bridges and their degradation could avoid heavy-handed and costly interventions.

The results of this project will benefit both engineering and statistics researchers and industry by improving the understanding the instrumented infrastructure and by taking the first step toward adaptive sequential control of systems.


In this project, we are currently investigating a prototype railway bridge instrumented with ~130 fibre-optic sensors that record changes in wavelength over time.

These time series of wavelength readings are converted to strain measurements that are used to monitor how the load of a train is distributed across the entire bridge. The sensor readings are correlated and exhibit time-lagged and dampened features due to the different locations of the sensors and the force exerted by an incoming train.  Structural engineers are interested in characterising these correlations as a train passes over the bridge in order to understand the strength and stiffness at different locations and also detect any changes in the bridges structure.


Animation of wavelength readings for sensors located on the girders of the bridge. Green denotes small strain where red and blue represent large strains that coincide with a train passing over the bridge.


Animation of predicted wavelength readings over the entire surface of the girder.

We are currently investigating various statistical methods to model the wavelength readings, that encapsulate the correlation structure between sensors. In order to detect degradation of the bridge, we shall monitor this correlation signature of the bridge and signal when there is a departure. This may involve using statistical techniques used in the change-point detection literature.






Further, this model will be used to predict the strain of the bridge at non-sensor locations and at future times. These models need to be able produce predictions along with a measure of uncertainty i.e. a confidence interval. This is of key importance to structural engineers, who always prefer conservative predictions.

After these statistical methods for modelling the sensor readings and detecting changes in the structure have been established, we will focus on how to deploy such methods across multiple instrumented bridges. This will allow us to monitor the structural health of similar bridges over time and perhaps indicate which bridges at most at risk. Industry partners i.e. railway companies who own and maintain such bridges, would be interested in using these data-driven techniques to guide inspection policies.

Part of the Alan Turing Institute-Lloyd’s Register Foundation Programme for Data-Centric Engineering.

Funded by:

lrf logo (002)