General Electric is embracing the concept of the “digital twin”—a data model of a specific physical asset—in a bid to eliminate unplanned downtime of aircraft engines and other systems. The initiative is part of GE’s drive to become a digital manufacturing company and builds on its new Predix software platform for the Industrial Internet.

A relatively recent concept, “digital twin is both a reality today and it is coming as we keep adding more product lines, engines and failure modes,” says Mohamed Ali, general manager of statistical performance and analytics for service engineering at GE Aviation.

Digital twin allows analytics-based maintenance. “The best we have today is condition-based maintenance. We boroscope an engine at intervals to see if it needs maintenance. Now we can use analytics to tell us how we are accumulating damage by specific part, by failure mode,” he says.

“Every time an aircraft takes off, we get data on ambient temperature, dust conditions, corrosive environment, weather conditions. We know the manufacturing data, the flaws in the part, its thickness. We know how it is being operated. So we know how it will accumulate damage,” says Ali.

“We develop expert-based criteria that tell us this amount of damage will exceed the limit. So we tell the customer to go look at the engine, and they will find damage that is at or close to the limit,” he says. “We have had several successful cases.”

The term “digital twin” conjures up an image of a virtual model of a complete engine, but GE is starting with life-limited parts and failure modes. “Over time, we may creep toward creating a complete digital twin, as the need emerges, but it will be on a cost-benefit basis—for us and the customer,” Ali says.

“Sometimes you never get any data because a part never fails. Do I need to model that part, as opposed to the limiting parts that cause the most pain? That debate is going on, but our goal is not a sexy-looking digital twin of an engine,” he says.

GE has started the digital-twin initiative with hot-gas-path parts in engines where failures can be costly—“in long-haul aircraft engines where you can get an expensive out-of-station failure,” he says. “It’s not as painful for short-haul engines, where all your aircraft are operating within the U.S.”

The process begins with data downloaded or downlinked from the aircraft and engine. These are cleaned and quality-inspected, then go into a “data lake” within the Predix cloud-based platform, which has been developed internally by GE Digital.

“That’s the beauty of Predix,” says Ali. “We do it once and leverage it many times. We give the analytics developers a clean set of data to run on Predix. The physics is the same, so they can develop an analytical tool for one engine, then leverage it across multiple engines.” 

Predix provides analytical microservices, or complex software applications, that are highly modular. “Copy and paste, and you have a microservice with a new coefficient for a different engine,” he says.

“We do it by engine serial number. Every one is tracked by failure mode on that part in that engine number—specific  mode, specific part, specific engine serial number,” Ali says. “It is all about understanding the variation. The average can be failure-free; it’s the variation that can cause an engine removal—one blade may be oxidized, but not the average blade.”

The rate at which models are run depends of the failure mode. “Some parts fail from low-cycle fatigue, and it takes a while to accumulate damage, so we may run the model every week or every month,” he says. “Others we run every night. We have not seen cases faster than that, but it may happen.”

Analytical tools are developed in collaboration with customers, because aircraft data are needed to validate the tools. Where the models are safety-related, the results are shared with all operators; otherwise it is by arrangement with each customer, Ali says.

Where the inherent physics is the same, GE tries to avoid creating customer-specific digital twins. “That works well when we have the data on how the engine is operated differently, in different environments. If we have a customer-specific model, it means we don’t fully understand the operation or environment,” he says.

The analytics focus of the digital-twin initiative “provides huge insight into how the engine is being operated, which can feed back into design,” says Ali. Over time, this may lead to more sensors being put on engines to collect more data, but there will be a balance, he says.

“I would love to have 25 more sensors, but we would not necessarily get the data we need, and we would add more opportunities for failure,” he says. “We still have opportunities to use the data we have. We are still in an early stage.” The engine manufacturer is working with GE Global Research on data “imputation,” algorithms and techniques to infer information when data is missing, he says.

So far, it is GE itself that runs the digital-twin data models. “We have a vision to allow customers to have the applications and get the results of the analytics,” Ali says. “It is still a vision. We are still in dialog with our customers. This may be a landslide change [for the industry], so we need to be very careful.”