How KLM Cityhopper saves costs and downtime.

Data Cloud

With predictive models for maintenance.

Aircraft maintenance is subject to strict laws and regulations. At regular intervals, the aircraft will be inspected and certain parts must, by definition, be replaced every few years. Mechanics are responsible for regular inspections and standard maintenance based on the age and usage of the aircraft. At KLM Cityhopper, however, they go one step further to make the current maintenance plan more effective and efficient. To purchase as efficiently as possible, schedule maintenance, and gain intelligent insights from varied aircraft maintenance history, they use artificial intelligence. More specifically, predictive data. With KLM Cityhopper predictive maintenance, they predict upcoming maintenance and by scheduling the aircraft in for maintenance, theycan prevent aircraft from being grounded for too long or needing more expensive repairs. This offers advantages for almost every industry where machines are in charge and downtime costs money immediately.

The maintenance team is in a hangar with working on the Embraer aircraft, the aircraft that KLM Cityhopper uses for short flights within Europe. A mechanic links a drive to the aircraft to transfer the so-called QAR or TCRF data (TCRF – Triggered and Continuous Recording Function), the information from the well-known ‘black box’. The sensitive data is split because not all data may be used. Those parts are then unlocked for processing. The ultimate goal of using that data is simple: to detect a upcoming failure as quickly as possible because the sooner a problem is detected, the sooner the aircraft can be fixed without unnecessary ground time andthe less it costs to solve. However, retrieving that data is only step one…

Making data accessible
“Getting the data is a complex process, but transferring the data with the newer Embraers does happen automatically,” emphasizes Wemerson Cesar. He works on behalf of LINKIT as a Data Engineer at KLM Cityhopper. Alyona Galyeva joined later when the team needed a Solutions Architect in addition to a Data Engineer. “The data from such a black box is endless,” she says. “It contains all the sensors’ information that the aircraft records during the flight.. The data is classified as sensitive, so not every parameter may be used and not every employee is allowed to work with this data. So Wemerson was the first data engineer to join KLM cityhopper because he had to make sure that the right data ended up in the right place before we could do anything with it.”

Building bridges on the data lake
As a data engineer, he says, Wemerson can not support the KLM team 24/7, but he does it daily. He ensures that KLM’s data scientists and business intelligence department provide the correct data for new insights and research. In short, he builds and maintains the pipelines that enable the flow of information. First, he connects them all to the data lake, the place where the current data from the aircraft, split and decoded, ends up. “Data scientists and business intelligence don’t have access to that,” says Wemerson. “My job is, among other things, to ensure that crazy or double values are removed, and the sensor data is enriched with, for example, flight data, such as times.” Next, that clean and enriched data moves to another place via a pipeline, after which it is accessible for a specific purpose, such as building a dashboard by someone from business intelligence.

Real-time data and the data scientist as a unicorn
It is quickly said that this data should preferably be viewed in real-time. “That is often unnecessary and not cost-effective,” says Alyona. It also applies to KLM Cityhopper. “It takes a lot of computing power, and it is of no use at many moments. For example, when an aircraft is flying, you cannot retrieve data, and you do not need it from an aircraft that is stationary in a hangar.” That is why the dashboards are updated every hour with the current information. Another thing we encounter in almost every case is the data scientists’ role, says Wemerson. “Often they are seen as people who know everything, when in reality they have to deal with statistics, training models, making predictions and maybe making a dashboard.” “It’s best to start a process like this with only data scientists,” says Alyona. “If you want to see whether AI is a solution for your issue in your company, for example. But once you know you’re headed that way? Then strengthen your team with data engineers! It prevents you from getting stuck at a later point.”

Challenges and Solutions
“Our role requires that we can also substantiate why something is necessary,” says Alyona. “That applies to every step, including setting up the environment and using the right tools. Working with AI is a paradigm shift.” She gives an example. “With traditional software, you monitor in real-time: does it work or does it not work? AI monitoring combines historical and real-time data, where you always have to compare. You are dealing with something continuously mutating, whereas you previously had to deal with something static. That requires a different approach, but you always have to be able to explain why.”

The value of migrating to the cloud also had to prove itself—especially given the sensitivity of the data. “Now we send the data to the cloud in a kind of gibberish, then it is translated there, the algorithm starts working with it, and we get a prediction, which we make illegible again and then retrieve it on location. There we convert it back to the actual data, and only then can it be transferred to a dashboard.” It’s a bit of work, Alyona admits, “but it takes a lot less effort and maintenance than doing everything on site. Plus, we keep it safe that way!” It also has to do with legislation: it influences what can and cannot be done with the data and the actual maintenance. “A model can predict that something will only wear out in eight years, but if the law says you have to replace it after four years, then you have to.”

Success on a small scale: that leaves you wanting more!
With ‘only’ eight people in the KLM Cityhopper team, the assignment was nevertheless manageable. That made for a relatively straightforward process. “But remember that AI solutions always take time,” says Alyona. “There are often different departments involved. It’s not just about building; it also needs to be tested and regularly verified by those who will use it. You must know if it works, not just in the short term.” KLM called in the help of LINKIT after two years. The project has now been going on for five years. But, once it works correctly, it pays off a lot. “The greatest added value of predictive maintenance is that you can already order scarce parts, for example, which reduces your downtime. Another example is scheduling the maintenance before the aircraft actually has a failure. This prevents unnecessary ground time and possibly a delay or cancellation for passengers, which improves your customer experience significantly,” says Alyona. The joint team first ran a pilot with oxygen and hydraulics data to discover the possibilities and to ensure everything worked correctly, which they later converted into actual models. These have now been fully transferred to KLM Cityhopper, which maintains the current models. “That’s the best thing,” says Alyona. “Once this turned out to be a success, other KLM departments also wanted to get started!”