Energy Speaker Series - Module 2: Digital Twins for Digital Transformation
Dr. Vili Panov, Siemens Energy
Prof. Diego Galar Pascual, Luleå University of Technology
Overview
Session 2.2: Gas Turbine Digital Twin for Performance Diagnostics and Optimization,
Dr. Vili Panov, Advisory Key Expert at Siemens Energy and visiting professor of Health Monitoring and Diagnostics in College of Science at University of Lincoln
This talk will discuss development of a Gas Turbine Performance Digital Twin by use of tools such as Simulink Real-Time™ and Simulink PLC Coder™. The developed Performance Digital Twin utilizes Real-Time high-speed computing and can be leveraged with various Enterprise and IOT Cloud Platforms. Proposed solutions are provided in a form of modular software architecture for a range of hardware platforms with corresponding functionalities to support model-based control strategies and advanced asset health management.
This project explored novel advanced techniques, which can meet the challenging requirements of increased reliability, improved efficiency and extended operational life of gas turbine assets. The Digital Twins based on real-time dynamic engine models has emerged as the most viable approach for solving challenging control and diagnostics requirements.
The developed real-time on-line Digital Twin technology has the ability to enhance current state-of -the-art offerings which are predominantly based on non-real time and off-line solutions. The devised solution highlights the next generation of Digital Twins that exploit modular functionalities distributed across the whole IoT chain consisting of Embedded, Edge and Cloud computational platforms. The gas turbine Performance Digital Twin has been deployed on the operational site and collected field data have been analyzed and presented in this study.
Session 2.3: Virtual commissioning for Wind Farms using Hybrid models: A digitization approach,
Prof. Diego Galar Pascual, Division of Operation and Maintenance Engineering at Luleå University of Technology and Principal researcher in Tecnalia (Spain), heading the Maintenance and Reliability research group within the Division of Industry and Transport
Digital twin is a virtual and digital representation of a physical entity or system. It involves connected “things” generating real-time data. That data is analyzed in the cloud and combined with other data related to the context around it. It can be then presented to operators and maintainers in a variety of roles, so they can remotely understand asset condition, history, etc.
Wind turbines are complex with respect to technology and operations that is why, a viable solution is to apply intelligent computerized systems, such as computerized control systems, or advanced monitoring and diagnostic systems. Indeed, the huge amount of information provided by wind farms convert this type of asset in the perfect candidates for digitization and deployment of digital twins.
However, the digital twin must be complemented besides the captured information in order to assess the overall condition of the whole fleet/system including the one from design and manufacturing which obviously contains the physical knowledge.
Therefore, the integration of asset information during the entire lifecycle is required to get an accurate health assessment of the whole system and determine the probability of a shutdown or slowdown avoiding black swans and other unexpected or unknown asset behaviors. Moreover, the lack of data on advanced degraded states due to early replacements makes the data-driven approach vulnerable to such situations. These hybrid models are expected to be shortly integrated as a part of the digitization in the digital twins created as digital mirrors of wind farms all over the world.
This talk will show the digital twins technologies for wind farms including hybrid models applying the maintenance analytics concept by the means of virtualization i.e. virtual commissioning of the assets through data fusion and integration from a systems perspective.
About the Presenters
Session 2.2
Dr. Vili Panov is a Chartered Engineer and Member of IMechE. Currently he is employed as an Advisory Key Expert by Siemens Energy Industrial Turbomachinery Ltd. He is appointed Visiting Professor of Health Monitoring and Diagnostics in College of Science at University of Lincoln. He received the Dipl.-Ing. degree in Mechanical Engineering in 1996 and the M.Sc. degree in Aeronautical Engineering in 2002, from University of Belgrade. Dr. Panov was awarded a Ph.D. in Engineering Mechanics form Cranfield University in 2006, and he has more than 20 years of experience in the turbomachinery sector.
Session 2.3
Prof. Diego Galar is a Full Professor of Condition Monitoring in the Division of Operation and Maintenance Engineering at LTU, Luleå University of Technology where he is coordinating several H2020 projects related to different aspects of cyber physical systems, Industry 4.0, IoT or Industrial AI and Big Data. He was also involved in the SKF UTC center located in Lulea focused on SMART bearings and also actively involved in national projects with the Swedish industry or funded by Swedish national agencies like Vinnova. He is also principal researcher in Tecnalia (Spain), heading the Maintenance and Reliability research group within the Division of Industry and Transport.
He has authored more than five hundred journal and conference papers, books and technical reports in the field of maintenance, working also as member of editorial boards, scientific committees and chairing international journals and conferences and actively participating in national and international committees for standardization and R&D in the topics of reliability and maintenance.
Recorded: 17 Nov 2021
Welcome to MathWorks Energy Speaker Series. My name is Neil Vili Panov I am Siemens Energy employee. And I'm also appointed visiting professor at the University of Lincoln. In today's session, I will talk about gas turbine digital twins for performance diagnostics and optimization.
In introduction, I'll touch on basics of digital twin technology. In section two, I'll talk about the basic building blocks and network connectivity of developed cyber physical system. In section three, I'll explain core digital twin configuration. And in sections four and five, I'll show some trial results. And I'll explain distribution of the system across various platforms such as embedded edge and cloud platforms.
As already mentioned in introduction, we will look at various concepts of digital twin technology. There are various definitions of digital twin concept in open literature. But for the purpose of this talk, we are quoting this one. A digital twin is defined as a virtual representation of physical asset enabled through data and simulators for real-time prediction, monitoring, control, and optimization of the asset for improved decision-making through the lifecycle of the asset and beyond.
Typically, we can recognize three different types of digital twins, namely product digital twin, production digital twin, and performance digital twin. Product digital twin is used for efficient design of new products. Production digital twin is used in manufacturing and production planning. Performance digital twin is used to capture, analyze, and act on operational data.
In this section, we will discuss overlapping nature of cyber physical and IoT domain. And I will present main building blocks of developed system. Developed cyber physical system was based on performance digital twin for small, industrial gas turbine.
This system contains virtual gas turbine as a part of cyber domain and physical gas turbine as a part of physical domain, which are closely integrated with controlled system. Extension of digital twin functionalities was achieved with connectivity to Internet of Thing and remote monitoring system platforms.
There are three distinctive domains within cyber physical system, namely physical domain, cyber domain, and IoT domain. Physical domain contains physical gas turbine unit, which is connected to the automation system via multiple sensors and actuators.
Cyber domain contains self-configuring virtual gas turbine, which enables gas turbines to be monitored and controlled by adaptation to external and internal health conditions. IoT domain is based on networked technologies, which offer seamless integration of data objects within information network.
In this section, we will look at configuration of core digital twin and digital twin build process. Develop performance digital twin contains five major functionalities, namely tracking, diagnostic, prognostics, optimization, and trip prevention. Tracking module accounts for engine-to-engine variation and engine degradation based on alignment of digital twin.
Diagnostic feature is used for diagnosing typical gas path degradation and fourth modes based on health parameters generated by digital twin. Prognostics module offers estimation of remaining useful life or gas path components based on regression modeling of health indices deduced by digital twin.
Optimization module offers performance optimization based on model-based control strategies utilizing digital twin virtual sensors. Trip prevention feature offers reduction of gas path-related trips based on analytical sensor redundancy provided by digital twin.
On this slide, we can see basic architecture of core digital twin. Core digital twin contains several blocks such as predictor, estimator, common tuner, and sensor diagnostic block. RTM predictor is a real-time, component-based, nonlinear, dynamic gas turbine model, which contains nominal health parameters.
Estimator is thermodynamic gas turbine evaluation model, which estimates actual components' efficiencies and flow capacities. Common tuner generates human health parameters based on smooth health parameters evaluated by estimator. Sensor diagnostic block is responsible for diagnostic isolation and accommodation of faulty sensors. This block enables self-configuration of virtual digital twin by reconfiguration of control system and digital twin tuner.
In this slide, we highlight main difference between gas turbine digital twin and the gas turbine simulator. Digital twin can run in multiple configurations, depending on health status of gas turbine sensors. Tuning of gas turbine digital twin is fully preserved in case when all sensors are healthy.
Another side of this configuration spectrum tuning a gas turbine digital twin is not active. Because combination of sensor faults removes confidence in estimation process. In case of gas turbine simulator, sensory information is not available. And hence, tuning of engine model is not possible.
In this project, we explore the continuous and discrete system implementation of development gas turbine performance digital twin. In continuous system implementation, we use continuous digital twin simulation model. This model was deployed on PC-based platform running real-time operating system.
In this way, we achieved hybrid integration. Continuous digital twin was integrated with discrete engine controller. Continuous digital twin was integrated with discrete PLC platform via PROFIBUS communication protocol. In discrete system integration, we use discrete digital twin simulation model.
This model was deployed on slave PLC. In this way, we achieve homogenic integration. Discrete digital twin was integrated with discrete engine controller. Communication between slave and master PLC platforms was achieved via native communication protocols.
For deployment of continuous and discrete solutions, we use dedicated tool chains. In case of continuous implementation, Simulink continuous digital twin solution was deployed on PC by use of Real-Time Simulink tool. In case of discrete implementation, Simulink discreet digital twin solution was deployed on PLC by use of Simulink PLC Coder tool.
Subsequently, generated software blocks were imported into integrated development environments, which are compatible with semantic and Allen-Bradley PLC platforms. And finally, discreet builds were deployed in two configurations, as a stand-alone solution deployed on slave PLC and as embedded solution deployed on master PLC.
Next section, we will see digital twin trial results, showcasing implemented model-based control and prognostic and health management functionalities. To illustrate performance tracking functionality, on this slide, we can see how performance digital twin can generate typically non-measured variables such as turbine entry temperature and compressor inlet mass flow.
On upper side of the slide, we can see how digital twin can generate those values across the typical speed sweep maneuver, which corresponds to different geometry settings within the compressor. For different VGV settings, digital twin was capable to adjust itself and align with the real engine and generate these nonmeasured variables.
On the bottom side of this slide, we can see how digital twin can be used to generate already measured variables such as compressor delivery temperature and compressor delivery pressure. In this example, we can see for the fairly fast transient maneuver, which corresponds to engine start up and run up, digital screen can generate those values. And we can see good agreement with measured values from engine.
In next slide, we can see examples of performance diagnostic functionality. In this example, we can see how healthy links generated by digital twin that corresponds to power turbine or gas turbine engine can be used for detection and isolation process. Healthy links that correspond to turbine component is compared with Both signature. And then model-based diagnostic agent can perform detection and isolation.
In this slide, we can see how model-based diagnostic agent can generate fault isolation index that corresponds to particular degradation mode of power turbine component. In this slide, we can see an example of almost prognostic functionality. Health parameters can be used for estimation of remaining-- for estimation of remaining useful life.
On this slide, we can see how health parameters related to compressor components, such as compressor efficiency index and compressor capacity index, can be used in this process. Regression model can estimate, what is the remaining useful life for this component in this case compressor, in respect of degradation mode related to power
Analytical redundancy functionality is demonstrated with compressor delivery pressure measurement. In this example, we can see how fault in compressor delivery pressure measurement can be detected, isolated, and accommodated by substitution of real measurement with analytical measurement. And later on, this virtual measurement can be used in different control strategies. In this example, we can see how engine can operate using virtual measurement during some low change maneuver and also during some very rapid transients.
And finally, in this slide, we can see how digital twin can be used for performance optimization or gas turbine acid. Digital twin can generate, typically, nonmeasured variables such as compressor inlet mass flow and combustor exit mass flow. And subsequently, those virtual measurements can be used for model-based control strategies, which can contribute to stability of the compressor and capacitor component.
In this final section, I will explain how we distributed and deployed digital twin functionalities onto various platforms such as embedded edge and cloud platforms. This development was carried out in several phases, where in first instance, we explored deployment of core digital functionalities onto edge PC-based platforms.
Performance real-time target machine was integrated. It distributed control system in the test bay, where we achieved point-to-point connection with this edge platform. Field trial was carried out using mobile real-time target machine, where we integrated this edge path from-- with custom and distributed control system. And also, we demonstrated remote connection with using our remote monitoring system facility.
In the next stage of this development, we explore deployment of digital twin functionalities onto PLC platforms. Our testbed configuration was based on semantic PLCs. And field trial was carried out using Allen-Bradley PLC platforms. We explored deployment of digital twin as a standalone application using slave PLC in both testbed and field trial configurations. But also, we deployed digital twin functionalities within the master PLC as embedded application.
In the next stage of this development, we explored the expansion of real-time functionalities embedded within edge and PLC platforms with known real-time functionalities encapsulated within various agents, where we deploy them on cloud and enterprise platforms.
We explored development and deployment of agents on mainstream platform, which is based on Siemens' IoT core. But also, we trialed development and deployment of various agents onto Mosaic platform, which is based on AWS or Amazon Web Services IoT core.
In this slide, we can see how in this project, we envisage so-called whole IoT chain. Real-time functionalities of core digital twin are deployed on edge and embedded platforms on engine and plant network level. These functionalities are related to engine tracking analytical redundancy and various model-based control strategies.
Expansion of this functionality is achieved by deployment of various agents on cloud and enterprise platforms. These agents contain functionalities related to diagnostic and prognostics, which, in return, can enhance asset optimization and various condition-based maintenance activities.
In summary, I would like to highlight some lessons learned from this development. This project reveals some challenges related to real-time dynamism and self-configuration, in respect to system and software complexity, ever-evolving system functionalities, and challenges related to robustness, safety, and security of developed system.
Integration of physical and virtual systems at multiple network levels offered some benefits in respect to seamless integration of heterogenic platforms, adaptability and scalability of digital twin functionalities, enhancement of physical gas turbine in the new capabilities. But also, we see this system is a neighbor of new products and services for gas turbine users.
So when we are talking about business and custom benefits, we are referring to improved robustness of assets, but also increased flexibility and operational cost reduction. Flexible deployment of digital twin functionalities distributed across various platforms, such as embedded edge, cloud, and remote system platforms, offered expansion of real-time functionalities of core digital twin with no real-time functionalities encapsulated within agents deployed onto IoT and RMS platforms dedicated to fleet and asset analytics.
With this, I would like to conclude this talk and thank you for your time. In case that you have some questions regarding this presentation, please do not hesitate to reach out using provided contact details. Thank you for watching this presentation.
Hello. My name is Diego Galar I'm a professor condition monitoring. I'm in Lulea University of Technology, and also Principal Researcher in Tecnalia for maintenance and reliability. And it's crucial for me to deliver, here in the Energy Speaking Series of MathWorks, this talk about real-time commissioning for wind farms using hybrid models.
Actually, what I'm going to present is different applications of industrial AI and how is these applications in energy can be powered with the support of some systems. In this case, also my own experience with MATLAB in the digital transformation of several companies.
And let me start by saying that in the case of where-- in industrial AI, we are used to work with artificial intelligence for very long. We have been working in-- during the long-- during the entire lifecycle of the asset designer manufacturing, during the operational maintenance, et cetera.
But now when we come to an industrial AI, it's something different. It's not the AI that you-- we use in the different domains like-- or purchased by internet, or Google, or wherever. We are doing customer profiling, prediction, and these services. But we are working in services for the assets, the assets that, basically, are grouped in different fleets and different components in different taxonomies.
And even though in-between the AI and the industrial AI, there are some similarities, we are playing with different pipelines. It's basically, in the normal AI for humans, we have very small amount of information, but very small amount of data, very rich in information. In assets, we have a huge amount of data. But it's very poor in information.
That is why analytics is more relevant. And that is why we have to move from the traditional descriptive analytics that mainly is the most common that we have been working, the oldest one, up to the cognitive analytics. That is where the AI is playing the maximal role.
Actually, when we go for descriptive analytics what we are doing, basically, is trying to see the abnormal things respect to the normal things. And this is why when we have some data sets, basically, what we want to see is the difference between healthy stuff and unhealthy stuff.
And this is what we need with the descriptive analytics. We have some data set. Something is normal. And something is not normal. This anomaly detection is extremely important, even though it's not a failure of detection. Because they are-- normally, we don't know if it's a failure or it's a false positive.
And in this regard, and the wind power is essential to do that. Because we have a lot of readings coming from the sensors. In this case, let me say that, for example, the preprocessing that we have to do from the data is extremely relevant. And this is very common, for example, when we analyze the gearboxes of the wind powers.
Actually, one example that I think is worth to mention is the TSA. The time synchronous average that they mapped up is well-established and deployed. It's, I'm using these signals, vibration signals, up-- using the time synchronous averaging and checking in which part of the gearbox I would have in the failure.
I must say that in this case, this is a good example of preprocessing and checking the anomaly with this TSA model that is deployed in MATLAB and has been very successful with some of the customers that I am working. When I move to the next step, diagnostic analytics, that is one step ahead. I am not saying that there's an anomaly in one gear. That is what we could see in the previous slides.
Here what we are doing is trying to identify not only that there are some anomalies, but also identify, what is the nature of this anomaly? And this is very important. Because we are identifying the failure. We do not say that this is abnormal anymore. We say that this is abnormal. And we take this as a failure.
And in this case, for example, in energy systems is, for us, very essential, the clustering and classifying of the failures. That we can-- I'm showing they're a cluster of different failures with different subcomponents in-- actually, in our HPAC system for cooling-- heating.
And we can see that in this case, using MATLAB, for the end customer, we have the chance to have a real cluster with the different failures and see which failures are the right ones, in order to be mitigated and be compensated with the losses that we may incur with the diagnosis.
Then this is essential. Because we are doing the clustering after the anomaly. And then we can identify and-- after the detection, which failure is going? But the diagnosis is not enough. When we do the diagnostics, we have the information, what is going on? And we have been able to detect it.
But we need to know, OK, we have detected that something is wrong with it. We identified what is going wrong. But we need to know how it's evolving. And this is the prediction. Predictive analytics is essential to see how the failure is moving from point a to point b. Point a, healthy status. Point b is where the asset is going to lose the function.
And predictive analytics is what we call the whole estimation, meaning useful life estimation, or also known as a prognosis. And in this case, also, it's very relevant when we do that. Because the common belief is that the prognosis-- so the remaining useful life-- is unique. You take one subsystem or one component. And this subsystem has remaining useful life.
And this is not true. We have to go for prescriptive analytics. What do I mean for prescriptive analytics? Actually, prescriptive analytics has the principle that different behaviors degrade the machine in a different way. In few words, when you have a machine working with different operational profiles, you have different degradation rates and, therefore, different prognostics.
And in this case, for example, if you go to the wind power, you consider that, certainly, when we go to and we check the performance of a wind, power, power versus wind speed-- and this is also a cluster with MATLAB-- you can see that certainly, certainly, the behavior of the performance of the asset is different, but also the degradation.
And this is essential. Because when we check-- for example, according to the standards, the wind intensity versus the wind speed. In order to check the turbulences and thresholds, et cetera-- if needed, to shut down my machine-- is essential to see that with different operational profiles, with different wind, with different type of turbulences, the degradation of my machine is going to be different.
That's why we have a fleet of assets that is behaving differently in function of different operational conditions. And that is why prescription is ahead of prediction, in the sense that this prediction-- in different operational scenarios and using the different regressions that we have in MATLAB-- definitely, we are able to predict very well where the conditions are going to be.
Worse for the machine. And then the machine maybe needs to be shut down in order to protect the asset integrity. Same for some of the subsystems. When we are working with the hydraulic group performance versus the power, it's the same. Actually, it's the power of the machine that is delivering. It's something that, of course, is rather related to that performance.
But some of the scenarios may be more harmful for some of the subsystems. And this is extremely important. That when we do some predictive maintenance or prescriptive maintenance to predict the remaining useful life, we should be aware that we have to consider the different scenarios. Because different scenarios, they will have different remaining useful lives.
And this is something that is moving towards the next step, that is, the cognitive analytics. What is cognitive analytics? It's when I'm able to take these decisions, these scenarios, but not with one person in front of the computer taking the decision if the windmill should be shut down or not in function of the turbulences. But these decisions should be taken automatically with no people there.
Actually, let me say that this cognition is essential nowadays. Cognitive maintenance, I think, is the top of the shelf in terms of maintenance power by AI, and especially after the COVID-19 coronavirus has demanded a number of special services for remote located and other related assets that need these decisions to be taken in a dramatic way. Because the humans, they have proved that they are the most reliable part of the system.
And in this case, of course, when we go to that step, we are in the highest level of AI for maintenance. But still, we need to do predictive reminders. Of course, it's, we are not in the point of cognitive analytics. And we have to move towards this cognition. And the first step that we have to consider is that we are moving towards PHM, the prognostic health management.
PHM is when we are looking-- the future, we want to foresee what is going to happen doing prognostics. And we are not happy just with diagnostics. For many years, we have been happy just looking back in the mirror. But what's happening with the diagnostics, that is OK. This failure is there. How we can mitigate this failure?
This is not enough. We have to be aware that we have to move to the prognostics. And if will look at the evolution of mankind, we can see that the evolution of maintenance is moving from corrective maintenance to preventive, then RCM to rationalize the maintenance, then see the end. Maybe, these are just checking with some thresholds.
And finally, finally, PHM, that is the maintenance based on the future happenings on the asset. And this is maybe the step before the cognition, with tailor-made predictions for the different components of systems and enterprises. Then if we want to offer good services for our assets, the method that nowadays seems to be the most flexible method to use as a service provider is the digital twin.
Let's try to twin the reality. Let's try to have some digital entity. And then this digital entity should provide these maintenance services. The digital twins are not new in this aspect, as we have been working with digital twins for very long. We have been integrating different types of information. But now that we have the different sensors and different systems already connected together with the domain knowledge, we can talk about a real digital twin.
A real digital twin, that if we look at their progress in different companies, definitively is very solid. Some of the companies in the energy sector are very mature for that. Also, in robotics, some of the leading companies like Tetra Pak in Sweden, that is very-- it's a leading company in the food sector doing the data analytics.
But we can see that the maturity is different. It's still not all the companies are in the same point of the evolution of the Digital Twin But definitely, the digital twin is there in most of the industries. And we can talk about the digital twin as a service provider, a service provider where we input all the sensor information and the domain knowledge.
And the output is all the services in terms of diagnosis and prognosis, and maybe additional services like spare parts optimization, is there troubleshooting, et cetera? And this is extremely relevant in some sectors. I would say that especially in energy, the wind farms are really, really interested in the Digital Twinning training for one reason that I commented, especially after the coronavirus.
We need to have good analytics in order to provide services, maintenance services, for the digital-- for the wind turbines, especially the offshore ones. Because they are remotely located. And they are unattended. And among assets, then we need to provide things. Because maintenance actions are extremely expensive. That is why wind farms are very suitable for this thing.
But then what is what we want from the digital twins? The digital twin, for me, is always-- I always say this-- is always The Picture of Dorian Gray. If you remember the book of the-- or the movie, is that this guy had some deal with the evil. That's OK. The guy was going to be forever young. And the aging effect was going to be visible in the picture. That, of course, most did not come to a point that anyone could see the picture.
This is what we want for our assets. We want the asset forever young and see the aging effect in the picture. And in this case, the common problem that we have is we try to predict the future of our assets based on the past of our assets. And this is a very, very common mistake. Because when we try to make a digital twin and we try to do such prediction, we are making a huge mistake. Because we are not considering that many new mechanisms of degradation are coming.
Of course, in the next years, is there-- we are going to have a number of digital twins powered by many technologies. Especially IoT is going to be there, powering-- providing more information from the field. But it still-- we are missing something in our digital twins.
What are the basic components of a digital twin? A digital twin needs a connected product. We need sensors. We need a platform where all the readings of these sensors are preprocessed and the analytics is running. And finally, we need some services that are provided by this platform.
If we have these three components, we will be able to have this digital replica that is able to do this forecasting. And then we will have this Dorian Gray picture, in order to have the aging effect in the asset before this happens in reality. Many times, people is happy with the stochastic data team. That is, basically, in fact, you have a large population of windmills. You can take directly the probability, the density functions of that. And you have the PDF of this fleet.
And basically, what you have is the probability of a failure based on distribution, failure distributions. But you don't have a real digital twin one-to-one. This is what we call a stochastic digital twin. And it's very similar to the traditional REMS analysis. But if you have one-to-one, it means that your every member of the fleet in the real life plus one digital replica, then we have a real-time model. It's a one-to-one with some unique relation.
And in this regard, in windmills, this is essential to provide what we talk in terms of IDMS Because at the end of the day, we cannot neglect that what we intend with this is forecast the failures, and then provide the logistic support to maximize the performance and extend the lifespan of our assets.
Then what digital twins we are going to have? It's a very common digital twin 1.0, where we have our digital twin with just the digital twin as some essential data coming from the field, PLC data sensor readings from the commission monitoring, et cetera. And we put all this data flow in the machine learning stuff. And then we detect some deviation of normality. Or we can do some forecasting.
This is enough. This sometimes is OK if you have really steady conditions. Some people in the windmills, they are doing that. These are in the wind farms. Remember that when you have this data doing 1.0, it's where you may be happy. Because every 10 minutes, discovers are providing a bunch of data.
And if you are in very, very strict conditions and your conditions don't change that much, maybe the twin based purely in operational technologies is good enough. Because if your conditions are not going to change that much, then you may be on top. But if we move to the convergence with IT systems, where we have something else, we don't have just the data coming from the field in the way of sensor ratings, discover data, et cetera.
If we have the need to fuse with taxonomies and ontologies, then we have something that is entirely different. When we want to reproduce what is a semantic ontology, it means that graph theory-- that, by the way, is with some companies, Spanish companies. Especially here for wind farms, we have been very successful deploying these graphs with the MATLAB. Doing the technology-- the semantic ontology in order to recreate and co-create the reality of windmill
What does it mean? It's to merge the events that you have in the CMMS together with the sensor ratings. Basically, if you have a
vibration problem, this vibration problem may trigger a shutdown. And this shutdown may trigger the purchase of a work order and the airfair of a spare part
And all these actions are related in the way of graphs, for example, in one ontology. And this is the relation between events and sensor ratings. And this is essential, especially when we are struggling with assets like the wind powers. In few words, what we need is to move from the transaction layer-- where we have a number of data-- to the content layer to co-create a solution in order to have the maintenance reality.
And this is what we call, basically, the IT and OT convergence. The IT and OT convergence is something that is very common. When we merge the information of the sensor ratings together with the events that are in the IT systems, like CMMS, ERP, et cetera, this is what I really like to call the digital twin 2.0.
And the digital twin 2.0, in this regard, is something that we are creating metadata. Metadata means that the essential ratings, they have some tax regarding the events that are happening. And then we find the explanation for some fortunate or unfortunate events that may modify the behavior of our-- a asset.
That is why if a sensor rating is dramatically changing, if you have the IT and OT convergence in the data doing 2.0, you will not have a false positive. But in the 1.0, a dramatic change in the sensor rating automatically is going to trigger a false positive. But the windmills are working in different contextual scenarios. We have a lot of data coming from the field in terms of weather forecast, et cetera.
All this data coming from the field, a company related to the asset, basically, is what we call context awareness What does it mean, context awareness Same asset, same wind turbine in the Norwegian Sea, or in Spain, or in China. The same turbine is behaving differently.
Then the sensor ratings. When processed, the AI should be able to see how the context is performing, what is the context that we are playing, and then make the prediction on the prescription based on the context that we are performing our function. And this is-- what I'd like to say there-- now I would like to say that this-- the digital twin 2.x. Because it's the metadata of the digital twin 2.0, but with more information. It's generated by the context information.
And this provides much better behavior when we are talking about same assets in different locations. But still, we have a huge problem. With digital twins, you can see what is going on. But you cannot do a proper forecasting. Why you cannot do the proper forecasting? Because still, still, we don't have information of the future mechanism of degradation.
And when we come to this point, we need, we need some confidence in these systems. Because many times, the maintainers tell me, OK, is it-- can we trust on these systems? Our confidence is limited for prognosis in the predictions that we are doing with AI for the maintenance systems.
Then how can I overcome this situation? If I'm starting the data in the blue area-- because so far, what I have is the ratings in the blue area. But I do not know how the asset is going to behave in the red area. Then how can I make any prediction? I make sure that the mechanism of degradation of the blue area are the same of the red area.
This is what I say, that you cannot predict their future based on the past. Then something is missing. The data-driven methods, maybe, it's not enough. Maybe I need to go create something of the domain knowledge together with the data-driven methodologies. That's why data scientists, they have to co-create a solution with the operational maintenance guys, bringing together the physics of the failure and the data-driven solutions.
And this is what we need to do. Because if we predict just based on the past, we are going to have a huge uncertainty. But if we use the physics of the framework together with the data-driven methods and we complement the data set with something else that I will comment now, then we will reduce the uncertainty. And we will be able to make one accurate prediction.
Then there is no way. We have to combine the data-driven with the model-based in order to have some hybrid, some hybrid models that can put some light on our system. Because remember that when you have data collected from the field, you are going to be able to detect and predict what is in your data, nothing else.
If a failure never happened and it's not in your data records, you will not be able to predict that failure. That's why the data science is very promising when you have all the data in your records. If you have all the data in your records, all the failure modes of your FMEA makeup, of your failure mode to affect analysis, they welcome.
Use data science. Because you will be able to predict everything. But if you don't have that information, you cannot make that prediction. Then the data-driven methods, they fit very well when you have all the failures that might happen. Or at least the most critical ones. But when you don't have that, you need to go for physical-based methods, where you can combine the power of the data-driven together with the physical-- with the physics of the failure.
Then the data science is not the solution for the prognostic health management. We need the data science to do the analysis of the data. But definitely, we need to find something to increase, to increase the dimension of these incomplete data sets. And once we increase the dimensionality of these data sets, then we will be able to perform data science.
And the hybrid models is something that combines that. The hybrid model combines the power of the AI together with the power of the physics of the failure. And the way to do these hybrid models is what I like to say, the digital twin 3.0. That is IT and OT convergence sensors and events together with the engineering technology. That is the domain knowledge, the physics, of the failure.
And the way to build this twin is very simple. You have to check which asset you have. Do you have a makeup? Do you tailor-make a failure mode effect analysis? And if out of 100 failures, you have just in your data records 10 or 20, remember that you have 80% of failures that may happen. And you don't have any record. Then you have to articulate failure of physics.
And how you can do that? Definitely, the steps that I'm following with my progress is as follows. One, I do my FMEA makeup. I define the places, the failure modes that are missing. And I check that, OK, these four failure modes, I don't have any record of these four failure modes.
Then in case these failures pop up, I will not be able to identify that. What should I do? Then let's create synthetic data to mitigate these. And what is this synthetic data? Synthetic data is data of this failure, but created by a physics of the failure or a physical model.
And in this regard, let me put two examples. One is for a component, a bearing. A bearing that this model is not that new. Let me say that I did with MATLAB with her multibody model. I'm doing past contact on EHL theory is-- and with this, I was able, a few years ago, to do the model of the bearing and be able to create synthetic data of different failures in order to detect these failures when these happened.
I must say that now in MATLAB, you have the model of the bearing in same scale. And this is really, really good scenario when you don't have to do what I did some years ago with MATLAB. Because now the component itself is there. And it's much, much easier.
But when I did that, is I was able to create the model to create different failures, synthetic failures. In this case, you can see the failure in the bearings with different speeds. The failure, the spectrum that I get, the synthetic one and the real one, they are quite similar.
The frequencies that I am getting, the failure frequencies are quite OK. And the feature structure that I can do from the synthetic spectrum is extremely good in order to do detection when this failure that never happened in my system may pop up. And then this synthetic data, what I do is I match the synthetic data with the sensor readings.
And once I match the synthetic data with the essential ratings, I have a bunch of data, much more complete real data and synthetic data that complete most of the landscape of potential failures. And I can do-- in that white box that you have in the middle, I can do pure data-driven techniques with the data sets that are not just real anymore, are a combination of real and synthetics. And then I can perform detection, identification, and localization, and, of course, prediction of their failures.
And when I do the data fusion, I merge the real data together with the synthetic data. And I perform the prediction, the prediction, the diagnostics and prognostics of my system. The good thing is that in Europe, there is-- I have around 90% of total. And false positives, false negatives are really, really minimalist now.
Then what I mean is that completing my data set with synthetic data created in this case-- totally in the MATLAB for this-- is essential to reduce the uncertainty and, of course, reduce dramatically the number of false positives. But this is for a component. But very recently in AI, in one of my projects, I have been able to create one Simscape model, Simulink model for HPAC, creating synthetic data of a number of failures.
And this synthetic data-- that is for a complex model, where we have compressor, the exchanger, et cetera-- you can see there that is a complex system. It's not a bearing. And this complex system-- sorry-- totally done in MATLAB is really interesting for me to create synthetic data.
This synthetic data do a different picture instruction and identify the feature instructions that are good enough in order to do the competition of the data sets that they have. Basically, in systems like one HPAC, it's-- many times, we don't have a lot of failure data.
And with this, what we get that is really interesting is we get a number of synthetic data that, eventually, when the failure pops up, it's going to come. But also, the feature that I am starting from that is-- they are totally transparent and simplest for my machine learning. And I am able to identify the features in the ranking. That is, they are the most powerful one.
In this case, what I can do is to virtualize my energy assets. And this is essential. And-- but what I want to give their-- the real-- for me, the real thoughts that I want to transmit to all of you is that the mutualization of the asset that is very nice looks very cute.
It's extremely interesting when you are able to merge the physics of the failure together with the data-driven. In this case, following the example that I show you with Simscape, you can see how I can predict-- in this case, in my Simscape, I can do the prognosis of my system.
But considering different life stages of my component, it means that I can make predictions in function of the degradation stage. And then the uncertainty is going to be entirely different. Then to reduce the uncertainty, I can have discrete, discrete life stages, and associate this with the component, and therefore make different, different predictions.
This is really, really good when I'm dealing with complex components. That, of course, the integration. What I did when I did several years ago, the models for companies like the bearing. Now in this case, with Simscape, for me, was crystal clear. That we could go for that integration. And definitively, really, really accurate and certain predictions.
But I still-- let me say that we have some challenges. The industrial AI should help me to see the invisible things as well. And this is maybe what we have to keep on working. We have the industrial AI. That is helping me a lot, as I have commented, in order to detect the failures and predict the failures that are happening according to my analysis. And I have a number-- limited-- of failures.
But still, there are many failures that I do not know about them. Because I never did the analysis of those failures. Therefore, maybe what I need to consider is that the industrial AI should help me with these predictions as well. And this is what we call the digital butterfly effect.
The industry AI should help me to see what is entirely hidden for me and is not visible even for me
or when I do brainstorming to consider which potential failures may pop up. And if I neglect some failure, maybe what I find in the future is a black swan.
Remember that a black swan is this element that is very rare. But what happens, the impact is very extreme. And also, when you look back in the mirror, you find an explanation for that. That's why, believe me that the black swan losses are still there. We need to be very clear in the black swan losses. And we should try to mitigate these losses. That's why it's important. It's important that we don't do everything according to normality. And we consider, also, the long turns and corner cases.
Definitely, the conclusion of this is the knowledge that is going to be together. It's combined by three things. On one hand, the data coming from the field, that is the orange box, is the data coming from your sensors. And you will do many things with that. If this data is complete and you have all the failures there, forget about the other boxes.
But this is not going to happen. You will never have more than 20% or 30% of the failures. Then you need the blue box. That is knowledge dimension, physics of the failure to complete the data sets that you don't have in the orange box. But even the orange and the blue box together, you always have some surprise. And your system should be robust and resilience to keep up with the surprise.
And this is what I like to say, that is the digital twin 4.0. Well, now even though we may have some surprises, the surprises can be incorporated to our core knowledge. And then this knowledge turns the black swan into white. And then the next time, next time this black swan pop up, it will be able to be identified by the system.
Then the digital twin 4.0, basically, core comprises the physics of the failure, the data-driven, the context-- because different assets in different locations behave differently-- and the surprises. If you consider these four dimensions, you will have an excellent digital twin 4.0.
Then as concluding remarks, let me say that the digital twins, when you go for the energy and especially for the wind power, you need unit hybrid models. Because you cannot make any prediction based on your data. Because, basically, in your data, you have just normality. And normality means no information.
When you take operational maintenance decisions based just on data variant you are taking your decisions based on 20% of the knowledge. If you do that, be aware that you may have a lot of false negatives and a number of increasing false positives. And this is not accepted-- never ever-- by the maintenance people.
In the wind power, we are trying to look now at powering on life extension. If you are having just data, you cannot predict, as I say, the future based on the past. Be aware that the life extension is just possible when you introduce, when you introduce, in this case, the physics of the failure.
And of course, we need real data integration. The convergence of OT, IT, and ET must be real. And we have-- in this case, I have some number of examples with MATLAB Simulink, Simscape that are really good, where this integration is possible. The moment that we have the degradation mechanism, and we can evolve this. And we can reduce the data.
Because the main goal for us is to reduce the number of sensors. And our AI should be able to work with a minimal number of sensors. The moment we can track evolutionary algorithms, then we will have the digital twin 5.0. Thank you very much. It has been a pleasure to be in this scenario speaking series by MathWorks. And I will be happy to answer any question. Thank you.