Digital Twins and AI: The Dynamic Duo

Under Pressure

?It’s 9:40 PM on a Friday evening. You’ve had a stressful month as the operations manager of your company’s largest facility, battling supply constraints and manufacturing bottlenecks, and now you’re looking forward to a long-awaited holiday weekend getaway with your partner. Back at the plant, things are humming along smoothly when an audible alert sounds in an operations control room, quickly followed by a graphic on the central screen showing an overpressure alert in a key part of the process. The lead operator snaps to attention, a knot forming in her stomach. If she shuts down the process, a large amount of valuable product will be ruined, and your company will miss an upcoming shipment. That’s the good news –?an unchecked overpressure could also result in a much worse outcome.

?

Thankfully, this isn’t the first time the operator has faced this scenario, though it is the first time she’s seen it in real life. Her training kicks in, and she turns her attention to the embedded artificial intelligence (AI) “Copilot” which has been monitoring the emergency in parallel. Within three seconds of the overpressure alert, the Copilot has evaluated multiple potential failure scenarios that could have triggered the alert. It has determined with more than 90% probability that a pressure transducer failure is to blame, based on the spatial-temporal pattern of inputs from other sensors in the process. In that brief time, it also retrieved 10 Hz signal data from a temporary cache and observed an instantaneous rapid pressure rise from one data point to the next and confirmed that such a rise is impossible based upon a fluid dynamics model of the system. Based on this data and analysis, the Copilot raises its confidence in a transducer failure to 99.9% certainty.

“Hey, Antoine,” the operator speaks into her handset, “I need you to check the pressure in the cooling loop ahead of the expansion valve on Line 3. I have an alert, which Copilot tells me is just a bad transducer.” Shortly after, Antoine confirms that the pressure is normal and that he will have the transducer replaced within 15 minutes. He completes the unscheduled maintenance and logs the repair in the plant maintenance system. Everything runs smoothly for the rest of the weekend, and the first you learn of the problem is a brief mention in the Tuesday morning operations stand-up meeting.

Thinking Ahead

As incredible as our fictional Copilot sounds, it’s entirely feasible using technology that is available today. The benefits are clear, but these benefits can’t be realized without answering two key questions:

1. How can I create and deploy such a system?
2. Is such a system economically viable?

Let’s start by addressing the first question. While creating and deploying a system such as Copilot is feasible, simply bolting on AI as a patch to an existing system isn’t the answer, and classical engineering methods are needed as well. You will need an appropriate system model that can evaluate different scenarios in both normal and abnormal operating regimes. The right tool for system modeling depends on the application, and while system models can be developed after the fact, they’re even more valuable when developed as part of the overall system design process. If you want to see examples of how such system models are built and their potential applications, MathWorks’ YouTube channel is one useful reference.

Once you have a system model in place, you will run various scenarios to simulate the behavior of the system and resulting sensor outputs. A key part of this process involves simulating failures and abnormal conditions, such as a sensor error, a material blockage or an actuator failure. Such scenarios are usually created as part of a Systems Failure Mode and Effects Analysis (SFMEA), which is an essential part of robust system design. By feeding simulation outputs into an appropriately constructed AI model, you can use machine learning (ML) techniques to train the AI Copilot to learn how the system operates across a wide spectrum of scenarios. Not only can a Copilot learn to interpret different sensor patterns and associated failure modes, but it can estimate the internal state of a system from these sensor values.

Of course, you also need people who understand how to use the system, and the digital twin of a system (the system model plus Copilot is a digital twin) is useful here too. Using the digital twin, simulated scenarios can be presented to an operator for training, in the same way that a pilot can practice in a flight simulator. Operator training and certification can even be integrated into operational processes and enforced using a learning management system (LMS) –?something I’ll cover in a future article. You need to work through human factors, something that a well-designed human-machine interface (HMI) can help with. The key is to expose people to these scenarios in a safe and controllable way, so that when they are faced with a real event, they know how to respond.

?

As you’ve probably deduced by this point, you will want to implement two digital twins. The first is a sandbox, isolated from the production system and used to safely evaluate different scenarios for training and development purposes. The second is a “live twin” which connects to the real-world system to provide real-time data, internal state estimates and diagnostics.

The second question is economic justification. Not every system needs a digital twin, but the more complex and expensive a system is, the more likely it is that a digital twin is not only desirable, but essential. Economic justification can be reduced to three simple questions:

1. How much does it cost to build and maintain the digital twin?
2. What is the expected economic value from avoiding downtime and other negative outcomes?
3. What is the expected economic value from optimizing system performance?

Although answering these questions is more complex than asking them, it’s not as difficult as you may think if you work with the right partners. You will need to make assumptions, but in many cases these assumptions can be refined using historic data from operations. The detailed engineering work needed to develop a system model usually needs to be done anyway, so that the incremental cost of building a system model can be less than you may think. Expected value calculations are probabilistic in nature but can be useful in estimating ranges of likely economic outcomes.

?

What’s Old is New Again

Although the ability to use symbolic modeling languages in combination with domain-specific, fine-tuned AI models is new, the idea of using a digital twin to handle an emergency is not. In his 2020 article Apollo 13: The First Digital Twin, Stephen Ferguson from Siemens PLM presents a compelling case that the first true digital twin was developed by NASA and that without it, the astronauts on that ill-fated mission would never have returned safely to Earth.

What has changed recently is the ability to use AI (specifically, certain machine learning techniques) in conjunction with digital twins. AI is applicable to these digital twins in two ways. As outlined above, a system model can be used to train an AI model to evaluate patterns of data from production systems, and then use those evaluations to draw inferences and make recommendations. AI can also be used to create reduced-order models of complex systems that operate in real-time. This is necessary for extremely complex systems for which classic physically based predictive models are too slow to be evaluated rapidly. If you’re technically inclined, this 2021 presentation by Dr. Karen E. Willcox from the University of Texas (Austin) provides an excellent overview of digital twins and the application of machine learning to create reduced-order models.

?

New Frontiers

The opportunity for combining digital twins and AI isn’t limited to managing manufacturing processes. For example, imagine a security system for a sensitive facility. AI monitoring can synthesize data from sensors in real-time, including video and audio feeds, to draw conclusions about intrusion risks and provide recommended actions.

Nor is this concept limited to precisely defined “closed” systems. One of the most interesting areas of development today is the use of virtual environments to train autonomous robots. The best-known example of this is the training of autonomous vehicles, but the core ideas are being applied to home robotics, underground mining, security monitoring and more. AI techniques figure prominently in these applications in ways that extend well beyond training ML models. Generative AI can be used to create diverse scenarios together with the ground-truth semantic data needed for unsupervised machine learning. New techniques are being developed to couple generative AI with procedural modeling to create the volume and variety of scenarios needed for robust machine learning.

Although I’ve painted an optimistic outlook for the benefits of combining digital twins and AI, challenges exist. A poorly scoped or poorly managed digital twin project can see cost overruns or disappointing performance, just as with any project. It’s common for vendors to overstate their capabilities in their zeal to win your business. If you use generative AI tools, you need to understand your legal rights associated with the underlying data used to train those tools.

If you are considering implementing a digital twin, with or without an AI copilot, I strongly advise you to seek professional advice. Learn more about digital twins and how I can help you achieve your goals at https://twinsightconsulting.com.

?

Note: Any mention of specific vendors or resources in this article is provided for reference only, and does not imply an endorsement of any vendor, product or individual.