GE Digital CTO on data, AI role in digital twin applications

Colin Parris has 22 years of experience as an executive at major divisions of IBM and now GE, but the CTO of GE Digital considers himself a data scientist first and foremost, and it shows. Ask him about the digital twin technology that is sweeping through product design, manufacturing and maintenance, and he'll recast it as a data management problem. Do you have the right data? If not, can you get it? How data-oriented is your culture?

Parris' data science skills, bachelor of engineering degree from Howard University, Ph.D. in electrical engineering from UC Berkeley and M.S. in management science from Stanford University helped prepare him for a career at the intersection of data analytics, AI and industrial IoT. He spent five years running GE's AI, analytics and data processing research group, where he led the initiative that produced more than 1.2 million digital twins for the company's manufacturing divisions and their customers.

"GE Research dealt with a whole bunch of rotating machines," Parris said, referring to the common denominator of the jet engines and the gas, steam and hydroelectric turbines the company manufactures. "What we do is take a business problem and put it into a data problem."

Colin Parris

In May, Parris started his new job as senior vice president and CTO of GE Digital, the subsidiary focused on industrial automation software and services. He characterizes his primary challenge at GE Digital as predicting things by using sparse data, such as simulation data and people's notes. His dual role is to push innovation in GE Digital's products so they provide value to customers, and to scale those products for GE's internal use.

Before coming to GE in 2014, he was a vice president at IBM, with stints involving Unix systems and the 6,600-developer division responsible for the systems software for IBM hardware, among others. Parris said then-GE CEO Jeff Immelt hired him to apply his IBM experience in the banking, communications and healthcare sectors to GE's major industries: aviation, transportation, mining, oil and gas, energy and healthcare.

In an interview, Parris discussed how to find the value in digital twins and the role that AI plays. The interview was edited for length and clarity.

How does GE use digital twin applications in its own manufacturing?

Colin Parris: A Digital twin, for me, is a living, learning representation of some form of asset or system. 'Living' means there is a continual stream of data that allows the model to be aware of the environment. There's also feedback data: 'When I do an action, what's the result?'

The other aspect is learning. I don't just want to learn from the sensor data from this one asset, I want to learn from things in the fleet, from simulators I built when I did the design. I also want to learn from people and look at all the logs the maintenance experts wrote. I want pictures.

It could be a part, full system, network or process. All of them are twins.

In GE, we use digital twins for three purposes.

The first is an early warning in a timeframe that makes sense for the business. The major thing is making assets as available as possible -- for instance, a jet engine. As the plane lands, it shows up at the gate and the pilots are conducting the usual pre-flight analysis. A light shows up on the cockpit. We are in failure. Why did that happen? The sensors only indicated it was a bearing failure. You can't do anything except offload the people from that plane.

But what would it be like to predict that bearing failure 30 days in advance? You need 30 days to find a way to get a different engine and plane to that gate at that time, and you need enough time to get the right routing for the crew.

The second is continuous prediction to predict when parts are going to fail. I need more time because in many cases these are just off-the-shelf parts. Building a jet engine and a good set of engine blades can take four to six months because I need to figure out inventory. If I have to stock all these parts, I need to go into the supply chain and get suppliers to produce this special nickel alloy and send it to 10 people before I can build this thing. If I can do continuous prediction by saying, 'Here's a level of damage on each part,' and I know six months in advance, I can optimize inventory so I know what I have to build. Otherwise I tie up hundreds of millions of dollars in inventory that is sitting there because I might need it.

The third thing we do is optimization, such as which plane to pull out so its engine can be repaired and put back in the fleet. If I can do that, I save a significant amount of money for you. If I can optimize fuel, if I can look at the way the pilots take off, how much fuel is used when it flies based on air currents, and then how much it uses when it lands, I can optimize the fuel.

The other thing you use digital twins for is simulations. You can do what-ifs.

Which digital twin applications have the best ROI?

Parris: You try to understand where the biggest gaps are that a customer has problems in. Where are they losing the most money? Where are they having the biggest risk, or where is the best revenue opportunity?

Then you think about the maturity level of the customer. Some are not digitally mature. They haven't collected the right set of data, or there are a lot of gaps in the data or it needs a lot of data processing.

Before, in my ideal world it was easy. I would show up and say, 'Let's digitize everything, and let's get master data management. That'll cost you $4 million.' Now, the discussion is: 'Let's go into the business, see where your biggest problem is right now, and then figure out what data I can collect to solve that problem.' If you have an $8 million problem, if the solution costs $10 million, you don't want to do it.

Next is probability. Do you want 90% accuracy that might cost $7 million, so you make a million-dollar profit, or do you want 60% accuracy that will only cost $2 million?

What challenges will companies face when implementing digital twin applications?

Parris: Companies usually have an MES [manufacturing execution system], a PLM [product lifecycle management] system and a variety of systems and say they collect a ton of data. But everybody usually collects data in a silo and for a specific purpose. If you know how databases and data schemas work, [you should first] define the question and then go after the data.

When the question changes and you have a dynamic environment, when you look at those systems and think they should give you the answers -- they were designed to ask a different question. You can't use the same data to answer a different question. You may have to add new data or change the data you collect.

Many people don't fully understand their problem. They're going from an industrial problem to a data problem.  

The second issue is: Now that I understand the problem enough, how do you collect that data? Have you been collecting the right data?

The third problem is: How do you go from a data problem to a data solution? Do you have the right data talent to do that?

What would you have done differently if you knew what you know now about digital twin applications?

Parris: I wish I had understood that digital transformation doesn't occur first. What occurs first is business process transformation. The minute that you begin to say there's a business problem I need to solve... and break up the problem in a certain way, it usually becomes about a value stream map.

Once I've done that, I can figure out where waste occurs, or where I can optimize better or produce new value. That's when I digitize. Those are the places where I can digitize the collection of information, the insights I can get and the actions I can take.

Once you do that, everybody unites. [If you can say] we're going to reduce or avoid costs, get more productivity, go out to new markets to get new revenue -- that's a rallying cry.

The second thing is an awareness of culture. I spent 20 years at IBM and live in this digital world myself, where every two to three years, a new sales book comes out or a new laptop. Every four to five years there's a new idea. It's always a revolution.

Then I come into the industrial world where a jet engine lasts 40 years, a steam turbine lasts 30 years and we have some that are 82 years old. It's a very different cultural experience. The people are trained to think first about safety and economic and process stability and high reliability.

All of a sudden, they're thrust into a world in which Google is buying electricity for data centers but it's also putting up wind farms and solar farms. These worlds that were perfectly so stable are now suddenly becoming very, very dynamic.

What is the role of AI in digital twin applications?

Parris: In a digital twin, you have an AI system working with a physical thing. I see them as inseparable.

It's entirely, in one case, a physics equation that says here's this temperature that's going to be in this location, or here's the strength of the metal. You have certain types of constants… that you believe are always the same in your system. The plane is going to take off a certain way -- I make that assumption.

The AI is working with the physics in some cases. In other cases, you don't have physical equations. When you look at a wind turbine and try to see how to characterize the wind that's coming into that turbine, you quickly realize that wind speed changes based upon the height of the turbine and is not a constant. The wind speed also changes at every hour of the day. That is such a complex thing it can't be modeled by a physical equation.

What we tend to try now is use AI and neural nets to model that. We may use AI to replace that entire equation. In many cases, we're taking physics equations and using AI to augment them.

The AI systems are part of the digital twin. So that combination of where we could use AI, where we don't have equations because the human mind can't figure out an equation of something as complex as the wind, or the AI is now adjusting what we assumed were always clear assumptions because of the variability of so many things.

What are the biggest challenges companies face in applying AI to digital twin applications?

Parris: First is the data. How do I collect the data? Because in many cases you're dealing with cultures in which the data is a byproduct. They use the data for something that is all focused on the machine itself. Collecting that data, understanding its value and cleaning it is a huge part.

The new industries -- the Googles and Amazons of the world -- started out as data collectors. Data is in their culture. One of the challenges we have is data has not been our culture in industrials.

The new industries -- the Googles and Amazons of the world -- started out as data collectors. Data is in their culture. One of the challenges we have is data has not been our culture in industrials.

The second challenge is: how do you get the right insights? Because you have a process that's running, and now I have to augment that process. I need data scientists who understand enough about the process to change it. And I need process engineers or mechanical engineers who understand enough about the AI capabilities so they can help direct the data scientists. You build that through experience.

The third thing is: How do you translate that in a way that enables business process transformation? Now I've got to put that insight inside your business process. I've got to find out how you make the decisions that impact your P&L [profit and loss].