Predictive Modelling in Complex Uncertain Environments: Optimised exploitation of physics and data

Lead Research Organisation: University of Sheffield
Department Name: Mechanical Engineering

Abstract

This project aims to improve the current techniques used to assess the condition and safety of offshore and aerospace structures.

The platforms used by the Oil and Gas industry in the North Sea were designed to operate for around 25 years in total. Over 600 of these platforms have now reached the end of their design life and the decision must be taken as to whether they can continue to be used safely or whether they should be decommissioned. For new offshore wind turbines, it is critical to have a good understanding of current structural condition so that maintenance can be planned optimally - unscheduled maintenance and downtime is extremely costly, owing to the difficulty of accessing these structures. Equally, in the aerospace industry, the ability to follow a condition-based maintenance strategy will save much time and money in avoiding unscheduled/emergency repair work.

This project brings together researchers from the University of Sheffield, who are experts in Structural Health Monitoring and nonlinear system modelling, with industry experts who are leading the way in the monitoring and assessment of offshore and aerospace structures. The aim of this collaboration is to develop the most accurate means possible of assessing structural condition using monitoring data.

The approach that will be taken here will combine the latest developments in artificial intelligence with more traditional methods that exploit understanding of the physics at work. Predictive models based on well-understood physics can often fall short of being able to explain complex behaviour, such as the loading an offshore structure will experience in a changing sea-state. This is where learning from measured data can be used to augment the model and improve prediction at times when the physics doesn't explain the behaviour captured by the sensors.

The combination of physics and data-based models will be used to improve the prediction of the forcing a structure experiences from a changing environment. An accurate quantification of this enables one to calculate the stresses a structure has undergone, which leads to a prediction of its current condition. A similar modelling approach will be used to help make predictions about the structure itself.

Finally, as well as improving the accuracy of the methods used to assess structural condition, the project aims to quantify the amount of uncertainty inherent in the models and algorithms that will be implemented. This approach acknowledges the fact that it is not always possible to make an accurate prediction of structural condition at a given time, but allows a confidence level to be assigned to each assessment made. To make responsible and optimal decisions concerning the repair or decommission of a structure, understanding the level of confidence one has in an assessment of structural condition is absolutely key.

Planned Impact

High-fidelity models of structures, informed by measurement data and expert domain knowledge, of the type proposed here, will ultimately allow operators to characterise performance, diagnose faults and deterioration and finally to predict how long the structures may continue to operate as designed, given current operating conditions. Although the application focus here is on offshore and aerospace structures, a new and general methodology is proposed here that will find applications across many industrial sectors, including chemical and process, ground transport and civil infrastructure.

Direct beneficiaries of the research here will be the industrial partners. The partners from the offshore sector are Ramboll (oil and gas extraction) and Siemens Gamesa (offshore wind). Safran Landing Systems (landing gear) and Dstl (military aircraft) are the partners from the aerospace sector.

In both of these sectors, enhanced predictive modelling capability is crucial for an efficient maintenance strategy. Reliable predictions of the dynamic response of structures enable a drive towards optimised condition-based maintenance. This is critical, particularly offshore, where structures operate in harsh and inaccessible environments; scheduled maintenance is consequentially costly and any unplanned maintenance extremely so (and can lead to large amounts of downtime). In the offshore wind context savings in operational costs can translate directly into reductions in the Levelised Cost of Electricity.

For offshore oil and gas, enabling a switch to condition-based maintenance is essential for lifetime extension. In the North Sea alone, over 600 platforms have exceeded their design lives. Where structures are believed to be undamaged and oil/gas resources persist, continued use is clearly desirable. A recent study estimates that reinforcing just 30 of these rigs in Denmark would cost 30 bn Danish Krone (DKK: approximately £3bn), while decommissioning would cost 15 bn DKK. Given that the UK has 360 rigs in the same situation in the North Sea, it is evident that extending the remaining life of the rigs is an issue of UK National Importance. The modelling capability developed here will inform and strengthen the decision-making process determining necessary actions towards the end of a structure's design life.

New, improved, wave force models will allow better control of uncertainty in load prediction, leading to more accurate estimation of fatigue accrual in structures. In the longer term and with wider applicability, any enhanced estimation of wave loading will inform the design of new offshore structures. This will be a major boon to offshore industry, as end users will be able to avoid the component overdesign that leads to economic and structural inefficiency.
The impact of the operational environment is equally difficult to quantify for aircraft; here enhanced modelling of structural response and, therefore, life assessment, is of equal importance. Access to this information allows extension of component use, therefore cost saving, and, importantly, the ability to schedule maintenance effectively. It has been estimated that predictive maintenance will increase aircraft availability by up to 35%, which would have major implications on the cost of air travel.

This fellowship paves the way to a new approach for predictive modelling. Beneficiaries in academia will be numerous, stretching from those in the SHM, structural nonlinearity and verification and validation research communities out to researchers working more generally in the modelling of dynamic processes. Dissemination will be achieved through conference presentations and publications, as well as at special-focus workshops hosted at the LVV. The data collected from the experimental campaigns will be made freely available, with wider dissemination impact achieved by developing YouTube videos of the large-scale tests.

Publications

10 25 50
publication icon
Deraemaeker A. (2019) Statistical analysis of damage indicators based on ultrasonic testing with embedded piezoelectric transducers in 9th ECCOMAS Thematic Conference on Smart Structures and Materials, SMART 2019

publication icon
Bull L (2021) Probabilistic Inference for Structural Health Monitoring: New Modes of Learning from Data in ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering

publication icon
Worden K (2020) On Digital Twins, Mirrors, and Virtualizations: Frameworks for Model Verification and Validation in ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part B: Mechanical Engineering

 
Description This project looked at trying to introduce physical insight into machine learning (AI) algorithms to try and improve how we monitor and manage our critical engineering structures. At the start of the project, there were very few methods and very little interest in doing this. Since that time, however, the research field has really grown and what we now call "physics-informed machine learning" has become a very vibrant area. The motivation for wanting to introduce physical reasoning or models into a data-driven learner were to try and counteract some of the problems we have when using engineering data to make decisions - the big one being that, although we might have lots of data, they will rarely cover all of the modes of operation of a system, or all of the environmental conditions that that structure will undergo. The hypothesis of the project was that if we introduced some physics into the algorithms that we use for tasks such as structural health monitoring, that we would be become less reliant on capturing data across the operational envelope of a structure - and this has been shown to be true.

The models/algorithms that we have developed in this project take into consideration a range of prior knowledge on how a system works/operates. If you are able to describe much of the behaviour of your structure or system then the residual models that we have employed are a good route to take; here the machine learner essentially works to account for unexplained behaviour in the physical model. If you know much less about your system, the constrained methods that we have looked at are a good starting point - these allow you to encode simple knowledge, such as the boundaries or shapes of a system. In the middle of these two approaches are hybrid methods where there is a strong interaction between data and physics model components - these models are highly expressive but there is still much to learn about how best to harness them. Work here is ongoing. For a more detailed introduction to the project, the book chapter on our webpage provides an overview and further links (https://drg-greybox.github.io/publications/).

In terms of results, the physics-informed machine learning approaches developed are providing models with improved predictive capability that are able to extrapolate (predict behaviours unseen in the algorithm training period). These models have shown to be successful for a number of structural health monitoring tasks including bridge deck deflection prediction for performance monitoring, wave loading prediction for offshore structure assessment and crack localisation in bearings (via acoustic emission). Again for more details and links to our open access publications please see our website - https://drg-greybox.github.io/.
Exploitation Route The models that we are developing are of use in other contexts within engineering and maybe applied in other disciplines, there are currently a number of funding proposals underway for the application of this technology for electric vehicles, personalised health care and manufacturing.
Sectors Aerospace

Defence and Marine

Energy

Healthcare

Manufacturing

including Industrial Biotechology

Transport

URL https://drg-greybox.github.io/
 
Description The fellowship started at a time when very few people were considering how physics may be embedded into machine learning algorithms and this was a particularly novel concept in my area - structural dynamics. I'm confident to say that my work, and the dissemination of it, contributed to the growth of what is now a very active field of research for many. In the last couple of years special sessions and workshops dedicated to what we would now call physics-informed machine learning (or something along those lines) have been appearing in our community and are well attended. Interest from industry is also increasing and indeed the findings of the fellowship are currently contributing to new technology development in the aerospace sector through an InnovateUK/ATI funded grant. This project started last year and we hope to be able to report impact in the future here as a result of it.
Sector Aerospace, Defence and Marine,Transport
 
Description Parliamentary breakfast
Geographic Reach National 
Policy Influence Type Participation in a guidance/advisory committee
 
Description Collaboration with Ramboll Energy 
Organisation Ramboll Group A/S
Country Denmark 
Sector Private 
PI Contribution The research team are working with Ramboll Energy to establish physics-informed machine learning models for predicting the health state of offshore structures
Collaborator Contribution Ramboll energy provide access to industry experts, access to data and models in current use. They have co-sponsored two PhD students on this project.
Impact See publication list particularly those with authors D.J. Pitchforth, S. Gibson and T.J. Rogers
Start Year 2019
 
Description Safran Landing Systems 
Organisation Safran Landing Systems, UK
Country United Kingdom 
Sector Private 
PI Contribution This project continues the collaboration between the PI in the Dynamics Research Group and Safran Landing systems. The research team are working to develop physics-informed machine learning models of landing gear for usage monitoring.
Collaborator Contribution Safran Landing systems are supporting this collaboration by offering their time and expertise to help guide the modelling work
Impact The work in this fellowship has lead to a new collaboration and project funded through InnovateUK and ATI.
Start Year 2019
 
Description Industry talks 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Industry/Business
Results and Impact The outcomes of the fellowship were presented and discussed in a number of forums with an industrial focus:

Invited talk "Physics-informed machine learning for Structural Health Monitoring." HBM Prenscia virtual Technology Day, November 2022.
Panel Member for Royal Academy of Engineering Technical Briefing on data-centric engineering. October 2022. Other panel members: Mark Girolami (FREng), David Lane (FREng) and Maja Pantic (FREng).
Invited Speaker at UK Magnetic Society day on Mechanical Engineering Aspects of Wind Turbine Design. Talk on "Physics-informed machine learning for wind turbine health monitoring." February 2022.
Invited talk "Grey-box modelling for Structural Health Monitoring." HBM Prenscia virtual Technology Day, November 2020.
Year(s) Of Engagement Activity 2020,2022
 
Description Invited talks 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact The outcomes of the fellowship were shared at a number of research focused invited talks around the UK and Europe, reaching a wide audience:

Invited talk on ``A spectrum of physics-informed machine learning approaches for structural health monitoring'', Tea-time talk at University of Cambridge, Engineering department, January 2023.
Invited talk on ``A spectrum of physics-informed machine learning approaches for structural health monitoring'', Workshop on Physics Enhancing Machine Learning in Applied Solid Mechanics Institute of Physics, London, 12 December 2022
Invited talk at Euromech colloquium on Uncertainty quantification in computational mechanics (618) on ``Probabilistic assessments of structural health with physics-informed machine learning." December 2021.
Invited talk on ``Physics-informed machine learning for wind turbine health monitoring" for the AURA CDT annual conference. Sept 2021.
Invited talk ``Physics-informed machine learning for problems in Structural Health Monitoring". Imperial College, London, March 2021.
Invited talk ``Physics-informed machine learning for structural health monitoring." Pipebots Programme Grant Seminar series.
Invited talk ``Grey-box models: Putting some physics back into machine learning models for structural health monitoring." The Unusual seminar series, an online series of seminars hosted by TU Delft, July 2020.
Invited Speaker at Strathclyde University: Guest seminar on Grey-box modelling for structural dynamics, November 2019.
Year(s) Of Engagement Activity 2019,2020,2021,2022,2023
 
Description Keynote presentations / talks / speeches on fellowship outcomes 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact The outcomes of the fellowship have been widely disseminated at a number of keynote talks to different communities, all of which have led to new collaborations, discussions and ideas:

Keynote speaker at ERNSI 2022, European Research Network on System Identification. "Physics-informed machine learning for structural dynamics in a Gaussian process framework'', Leuven, September 2022.
Keynote speaker at European Workshop on Structural Health Monitoring, "Physics-informed machine learning for Structural Health Monitoring'', 2022 (600 attendees).
Keynote speaker at 3rd Sheffield Workshop on Structural Dynamics, ``Physics-informed machine learning for structural dynamics'', Sheffield, December 2020.
Keynote speaker at 2nd Sheffield Workshop on Structural Dynamics, ``Combining machine learning and physics-based modelling for problems in structural dynamics'', Sheffield, November 2019.
Keynote lecture at Nonlinear Benchmarks Workshop, "Towards grey-box models for structural dynamics'', Eindhoven, April 2019.
Year(s) Of Engagement Activity 2019,2020,2021,2022
 
Description Lectures 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact A number of formal invited lectures were given on the topics in the fellowship, meeting a broad audience from a number of different disciplines:

Invited lectures and deep dive coordination for an Isaac Institute Programme on the mathematical and statistical foundation of future data-driven engineering, Spring 2023.
Invited lectures on "Physics-informed machine learning for structural health monitoring'', Summer school on "The era of AI and digitalization for structural applications'' at TU Delft, June 2022.
Invited lectures on "Suppression of Confounding Influences" as part of a Graduate course on Structural Health Monitoring, Politecnico di Milano, May 2021.
Year(s) Of Engagement Activity 2021,2022,2023
 
Description Physics informed machine learning workshop 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact Hosted a virtual workshop day on Physics-informed machine learning for structural dynamics. Event took place during the 3rd Sheffield Workshop on Structural Dynamics,
Sheffield, December 2020.
Year(s) Of Engagement Activity 2020
 
Description SYSID special session 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact Special Session at SYSID 2021. Together with T. Rogers and M. Schoukens (Eindhoven), we organised a special session at SYSID 2021 (IFAC) on physics-informed machine learning. This was one of best attended sessions of the whole conference.
Year(s) Of Engagement Activity 2021
 
Description Sheffield Society of Model & Experimental Engineers 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Regional
Primary Audience Industry/Business
Results and Impact This was an outreach talk given to members of the Sheffield Society of Model & Experimental Engineers. The topic was on the use of machine learning and artificial intelligence for assessing the health of structures. The talk was very well received and sparked lots of questions and interesting discussion. The talk will lead to a follow on engagement activity at the Laboratory for Verification and Validation, and has also forged a new link between the Dynamics Research Group and ANSYS.
Year(s) Of Engagement Activity 2019
 
Description Structural Dynamics Workshop 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact The 2nd Sheffield Workshop in Structural Dynamics was an academic workshop hosted in Sheffield by the Dynamics Research Group. It was themed around the EPSRC Fellowships of Keith Worden (EP/R003645/1) and Lizzy Cross (EP/S001565/1), with one day associated with each project (population-based SHM and grey-box modelling respectively). The third day focussed on V&V and digital twins and incorporated full-scale structural test demos at the newly-operational Laboratory for Verification and Validation (LVV.ac.uk).
Year(s) Of Engagement Activity 2019