Expert Q&A: Protecting national infrastructure assets

The ways in which infrastructure assets including wind turbines and bridges are monitored and maintained are set to be transformed in a ground-breaking project.

Led by Sheffield University and funded with a £7.7m EPSRC Programme Grant, the ROSEHIPS project brings together experts from academia and industry to solve the challenge of safely and economically safeguarding current and future infrastructure. ROSEHIPS (Revolutionising Operational Safety & Economy for High-value Infrastructure using Population-based SHM (Structural Health Monitoring)) runs until 2027 and involves Queen’s University Belfast, the Universities of Cambridge and Exeter.

Meet the experts

• Professor Elizabeth Cross, head of the Department of Mechanical Engineering, Sheffield University
• Professor Nikolaos Dervilis, Department of Mechanical Engineering, Sheffield University
• Dr David Hester, senior lecturer, School of Natural and Built Environment, Queen’s University Belfast
• Professor James Brownjohn, Professor of Structural Dynamics, Exeter University
• Professor Keith Worden, Department of Mechanical Engineering, Sheffield University

Why is ROSEHIPS necessary?

KW: A real challenge to monitoring large infrastructure is the lack of data from these structures, particularly damaged-state data. That said, even when there are data available on a given structure, it is not obvious whether they can be used to inform management i.e. provide decision support on repairs or improvements on another broadly-similar structure.

While we can use engineering judgement to identify potential similarities between structures, this is subjective and is likely to lack consistency between different analysts.

To overcome this problem, ROSEHIPS will develop and use a formal mathematical approach for calculating a similarity score between structures. This aim is achieved by representing structures in the form of mathematical graphs and then using graph theory to calculate a similarity score between them. For structures with a high similarity score, it is likely that data from one will be helpful in managing the other.  This systematic approach for comparison is a huge step in the field, as for the first time, it allows the limited data that are available to be leveraged, by exploiting them to manage other structures. 

Three road bridges come to mind: Orwell, Dartford, and Forth. How would ROSEHIPS be applied across structures such as these?

DH: One of the key concepts of ROSEHIPS is to model the structures of interest mathematically via a graph representation, then search the ROSEHIPS framework/database to find similar structures for which there are data available. For example, if you wanted to perform diagnostics on a box-girder bridge, such as Orwell, you would represent the bridge as a graph and then search the ROSEHIPs framework for other bridges that had a high similarity score relative to Orwell. Assuming you could find some, then you can potentially exploit data available from these other bridges to make inferences about Orwell. The same concept holds for Dartford [cable-stayed] and Forth [suspension]. So two key strengths of the ROSEHIPS framework are that: (i) it can act as a home for available data, and (ii) it is searchable in a systematic way to allow the analyst to find bridges that are relevant to the bridge of interest, in terms of decision support.

Speaking more broadly if we take suspension bridges as an example; while these are high-profile individual structures and often feature elaborate monitoring systems (such as those deployed on long-span bridges in the Far East – like Tsing Ma and Stonecutters), there are behavioural features that they share; for example, the behaviour of the joints and bearings at span ends. This possibility means that a relatively sparse SHM instrumentation package on one bridge can leverage behaviour observed on other members of the monitored population to identify actual or prospective anomalies. This strategy works better with smaller (not high-profile) bridges where highway operators find it difficult to argue a business case for extensive monitoring.

The fundamental application of machine learning methods will entail where structures are identified as similar but with some key differences - here we will be developing physics-informed and transfer learning methods to make comparisons possible.

For offshore, how would ROSHIPS be applied to a development such as Hornsea One with 174 wind turbines?

ND: This project introduces new ways of thinking about offshore wind farms and wind energy performance and health monitoring, as well as data analysis and processing. Until recently, the way the industry and the research community have tried to address the monitoring and performance problem has mainly been on an individual-structure or component basis. However, it is true to say that monitoring and performance analysis of wind turbines is still facing challenges on this individual basis because of a paucity of damage-state data, missing labels, operational and environmental fluctuations and/or computational costs. Consequently, population-based monitoring will bring a step change in methodologies able to address some of these challenges - from looking into whole wind farms as a population and different wind farms across the world as even bigger populations.

What sensor configurations do you have in mind?

JB: The sensor configurations are not prescriptive, the ROSEHIPS Framework will accept whatever sensor data have been collected on a given structure. For large cable-supported bridges, instrumentation tends to be similar and often links back to the South East Asian experience; these structures have tended to harvest vibration data using accelerometers. For smaller-scale structures, such as more common highway or rail bridges, the type of data collected will be more varied but could include several different modalities: vibration, acoustic, static/tilt etc. 

How will machine learning (ML) be applied in this project?

EC: To date there are very few examples of ML being deployed effectively for managing large in situ Civil Infrastructure and even fewer (if any), examples for populations of structural types. The question is therefore of how you can most effectively exploit and develop machine learning (or particular types, such as transfer learning), for managing large-scale civil infrastructure. Furthermore, the combination of machine learning/artificial intelligence with physics can offer a very novel and powerful approach, which will enable informative and robust predictions in the health-monitoring setting.  

Some would argue that the research will come to a homogenous/heterogeneous crossroads. Will ROSEHIPS diverge, or are you proposing a ‘one size fits all’ solution?

ND: A key strength of the ROSEHIPS framework is that, whether the structure of interest is from (1) a relatively homogeneous population (e.g. a wind farm where all the turbines are quite similar) or (2) a heterogeneous population (e.g. a highway bridge where the bridge is different to other bridges along the stretch of road); ROSEHIPS handles both situations in the same way.

Admittedly for (1) ROSEHIPS will likely find a greater number of similar structures, making it easier to find data/information relevant to the structure of interest. However, crucially, for the more-challenging case of (2); as the number of structures stored in the framework grows, it is more likely to find matches.

This ability to accommodate all structures in a single framework is an enormous step forward, as for the first time, it provides a mechanism to collate all available data and then exploit the data to maximum effect.

As the data in the framework grow, their power will also grow. In this regard there are some parallels with open-source computer vision libraries; as more data are available the power of pattern recognition methods improves. Again, the fundamental machine learning developed provides the key technology to bridge the gap between dissimilar structures.