Four blue planes and one white plane are in KLM’s maintenance hangars this afternoon. It’s a lot, they think here, a plane has to fly. In a few years, there may not be as many planes here at the same time. Maintenance can be made smarter with the help of artificial intelligence, shows ReMap, an EU-funded study by KLM, TU Delft and a number of other institutions and airlines.
The hangars smell of a garage. Green paths indicate the footpaths. A group of men on scaffolding scratches the body of an airplane. Another plane is on its way. A runway can be seen through the giant doorway, and the control tower in the distance.
Here are counted in flight hours. Depending on the type, an aircraft must land after 1,000 to 2,000 flight hours, after 200 to 700 times or 3 to 4 months – whichever occurs first – ‘in’ for an ‘A-check’. Then he stands here for 24 hours and the whole plane is inspected and grown.
“Time-based maintenance is not efficient. Time does not necessarily indicate whether an aircraft is healthy, “says Floris Freeman, who has done the research from KLM. Parts that are still in order are replaced regularly because the maintenance schedule says so. KLM would much rather perform maintenance based on the technical condition of an aircraft, condition-based maintenance (CBM) in aviation jargon. This prevents waste of time, material and manpower.
Another part of the maintenance is unplanned, which CBM is also a solution for. “Some things just have to break,” Freeman says. These are non-critical parts, such as an air conditioning system, of which there are several in an aircraft. “But the agreement is that such a thing must be repaired immediately, if one notices it. Whether a plane is in its home port of Amsterdam or in Paramaribo. ” It regularly leads to delays. In future, such maintenance must be performed just before the crash.
“Aircraft maintenance is about handling faults,” says Paul Chun, technology director for KLM’s maintenance department. “In the previous generation of aircraft, there was a lot more duplication. For example, they had four engines, if one broke you had three-quarters of the pressure left. Now the engines are so reliable that aircraft can handle two.” Chün sees condition-based maintenance as the next step.
“Soon a plane will only go to the hangar if it has to”, says Bruno Santos from TU Delft. Countless sensors then continuously monitor what is happening in the aircraft: temperature, pressure, vibrations, how much air is flowing through a pipe, etc. For a Boeing 787, a large long-haul aircraft, 165,000 parameters are involved.
“The next step is to understand what all this data means,” Santos says. If a sensor detects small deviations, an algorithm responds. In a meeting room in the hangar complex, Santos shows on a screen how the database has been translated into a ‘health index’ with colored dots.
“We could learn a lot from others,” Santos says. Also in health care, infrastructure and wind turbines, a lot of maintenance is carried out on the basis of technical condition. “They use similar sensors and similar algorithms there. The new thing is that we do so much at the same time and that we also plan maintenance on an algorithm-driven basis.”
“Also new is the IT platform we have developed,” says Freeman. “We were looking for a way to train algorithms with data from multiple airlines without having to share competitive sensor data.”
With 165,000 parameters about an aircraft, it would be said that KLM alone already has enough data available to train the algorithms. “It’s disappointing. In the aviation sector, safety comes first. So we have very little data on failing systems,” Freeman says. “It’s difficult.” If you want to learn an algorithm to recognize dogs and cats, and you train it with only pictures of dogs, then it can not designate cats very well. “The solution: the models are shared and updated once they have been trained with data from several airlines.
Test runs have been carried out over the past six months, so far only with KLM data and only on non-critical systems. “At least twice, we have prevented delays with a timely repair,” Freeman says. It has also been tested in the laboratory. Aircraft largely consist of structures made of composite, fiber-reinforced plastic. A problem with this material is that damage is often not visible on the surface. By mechanically shaking parts of a fuselage in the laboratory to mimic the conditions in the air, indicators based on sensor data for this have also been developed.
“Our models show that the total maintenance time can be reduced by 20 percent. It is one day a year, less by plane, ”says Santos. “In the long run, the weight of an aircraft can also drop by 10 percent.” The latter is due to the conservative nature of the industry, which always maintains large safety margins. If there is more reliable insight into the technical condition, fewer systems need to be duplicated and the construction can become even thinner.
But it is the music of the future. In fact, it will take until 2035 before condition-based maintenance is widely adopted, the European Aviation Research Advisory Council said in a press release on ReMap. “The planes we use today were designed 20 years ago,” Chun explains. “The cycle is long, it really takes a while before planes are equipped for this by default.”