The behavior of super-complex, multi-physical systems is difficult to predict. So how can you do thorough predictive maintenance? The High Tech Systems Center of Eindhoven University of Technology puts forward researchers who use artificial intelligence and digital twins to tackle this problem. A good dash of physics makes all the difference.
Dutch machine builders sell their high-tech systems all over the world. If one of the machines in the field fails, this can be very expensive. Depending on the application, a customer can look at a financial setback of a million dollars. Thus, machine builders are pushed by their clients to find solutions that allow them to accurately predict when their systems will fail. That way, the necessary repairs can be carried out or the (almost) worn parts can be replaced during a planned shutdown so that losses are kept to a minimum. How to properly estimate when a machine will malfunction is the field of predictive maintenance.
To some, maintenance may sound like a job for men in dirty overalls, but it actually requires good fundamental and scientific research. At Eindhoven University of Technology, the contribution of artificial intelligence to mechanical engineering is being stimulated, completely in line with the merger of the High Tech Systems Center with EAISI. A research project is underway at the university on system monitoring and predictive maintenance based on digital twins, with ASML, Canon Production Printing and VDL ETG as industrial partners.
“For simpler machines, it’s fairly straightforward to set up a model that you can use to analyze system behavior and identify potential errors,” says Carlos Murguia, assistant professor in TUE’s Dynamics & Control group. “This is much less evident for today’s extremely complex systems such as ASML’s EUV machines.”
Never enough data
Murguia explains that about thirty years ago, predictive maintenance was based on the use of so-called first-principles models. “With the Euler-Lagrange equations for motions and the physical laws for heat transfer and electromagnetism, among other things, you had enough physical baggage to build a model that could predict system behavior. You compared the outcome of that model with the sensor output of the machines in the field. When an inexplicable difference occurred, you knew something was wrong and you could send a mechanic to check it out.”
This approach worked very well when the machines were still physically manageable. “With the current multi-physical systems, it’s extremely difficult to describe all processes with those basic physical rules,” notes Murguia. “The various physical phenomena are simply too closely intertwined. And there’s an important new factor at play: software. Also, there are all kinds of communication phenomena that you have to take into account. It’s almost impossible to capture everything in a few simple equations.”
With the rise of AI at the turn of the century, companies chose a different route. “Because the physics had become too complex, they switched to artificial intelligence,” says Murguia. “Based on an enormous amount of data and with a neural network, you can also estimate the risk of malfunctions. A black-box approach, without a model. The problem is that you never have enough data. And you can’t give any guarantees about the result from the AI. When new, unknown data comes along, the system doesn’t know whether it’s an error or the situation simply hasn’t occurred before.”
No free hand
In recent years, methods that combine physics and artificial intelligence have been emerging. Murguia: “Although you can’t capture all complex interactions, you can certainly describe parts of the machine using, for example, transfer functions. And based on a kinematic model, you can put constraints to the movements of, let’s say, robot components relative to each other. You can combine those chunks of physics and AI into a hybrid model.” The result is a digital twin that’s tailor-made for predictive maintenance and allows you to make smarter predictions about when maintenance is required.
The success of this hybrid approach depends on the complexity of the machine. The more complicated the processes, the greater the possible gain. “For a simple robot manipulator, you can write out all the physics and the ‘old-fashioned’ approach with first principles will suffice. But for complex multi-physical machines like Canon’s production printers, the gain is substantial,” says Murguia, who finds it difficult to quantify the performance improvement. “It’s strongly determined by the application, but for the applications where we’ve used it now, cooperative driving with autonomous systems, we see approximately 10-15 percent better results.”
Currently, there are two ways to combine physics and AI. The first starts with the chunks of physics you can write down. Each of those pieces yields a small residual. “A residual is the difference between the output of the sensors and the model’s estimation,” explains Murguia. “With those residuals, you can train a neural network, so with historical data and with new data from the model. The new AI algorithm is then better trained and can make a better prediction.”
The second approach is to incorporate the physical equations into the AI algorithm. “How machine learning works is that a neural network tries to find the best fit for the input and output data. It doesn’t take physics into account at all,” says Murguia. “We’ve introduced those physical constraints in the learning part. We’re still training the algorithm to get the best parameters. However, we no longer give it a free hand but reduce the solution space by allowing only meaningful answers that comply with the physical rules.”
How many chunks of physics do you need to make a good prediction? According to Murguia, that isn’t the right question. “Our main goal is to see whether we can make it work. Can we add a number of physical rules to a purely data-driven AI twin and make a better prediction? The answer to that is – usually – yes.”
As a researcher within TUE’s Dynamics & Control group, Murguia focuses on dynamic models of mechanics. Other scientists within the group look at multi-physical models in which acoustics, electrostatics and thermal processes also play a role. All this domain knowledge is valuable to use in a digital twin.
Murguia emphasizes that there will always be a difference between physical theory and practice. Artificial intelligence can compensate for this. “The algorithm can learn that and properly characterize the various errors, provided there’s enough data available. Fortunately, this new approach requires much less data, perhaps only half compared to a pure black-box approach. After all, with the extra rules of physics, you give a boost to the neural network search.”
How does the hybrid approach translate to less complex systems?
“All modern machines are cyber-physical systems with a digital entity that receives signals and communicates with other components and networks. You can’t capture all of that accurately in physics. But if you can describe parts of the system with the laws of physics, you can benefit from using AI. First, you can incorporate the software and cyber-physical behaviors into the AI. And second, you can use machine learning to close the gap between model and practice. Even with a basic model, you can already monitor a machine very well because you learn how the two differ.”
How does the system know which discrepancies are also system errors?
“That’s step two. After you’ve abstracted the system behavior with a digital twin, it’s time for the decision algorithm. There will always be a difference between the real machine and its twin, but is it because of a system error or because of a modeling error? How well you can determine that depends on how accurately you’ve characterized the healthy behavior of a machine. So, that’s an important point for us. We try to describe errors as accurately as possible based on the sensor values that match such an error. The better you do that, the higher the monitoring performance.”
“It’s not all that black and white, by the way. We make a probabilistic statement about an error occurring. We can never be 100 percent sure until a mechanic does a physical check. Sometimes, an alarm bell will go off unjustly. In practice, you’ll have to make a business decision when you want to send that mechanic over. When the probability is above 40 percent? Or above 50 percent? Either way, our digital twin will be able to predict many of the system errors.”