Machine Learning in Adaptive Training for the Aerospace Industry

The modern aerospace industry generates very large amounts of data, mostly in the interest of safety and efficiency. Every commercial aircraft is equipped with a flight data recorder, also known as a ‘black box’ used to record information about every aspect of each flight. The recovery of the black box is one of the highest priorities in any incident investigation, and data retrieved is used to make the industry, as a whole, safer.

More recently, data scientists have gotten involved in aerospace to find possibilities for operational improvements buried in the data. GE, which manufactures aircraft engines, sponsored a competition at Kaggle, where data scientists competed to design new routing algorithms for flight planning that optimized fuel consumption and took into account variables such as weather patterns, wind, and airspace restraints. The winning routing algorithm was designed by a José Fonollosa, and showed a 12% improvement in efficiency over actual flight data.

Meanwhile, various technology startups are solving other tricky predictive problems using machine learning, such as more accurate arrival time estimation, or finding optimal takeoff parameters. While such operational improvements lead to cost savings for airlines, one major problem remains unsolved—the training of sufficient numbers of pilots and crew for tomorrow’s airlines.

Boeing estimates that in order to meet the projected demand for air travel, the industry will need to add 617,000 new pilots and a similar numbers of supporting crew by 2035. Much of this growth is being driven by Asia, where a new middle class is eager to enjoy air travel for the first time and airlines are meeting the demand by purchasing hundreds of new aircraft—but who will fly them?

As data scientists, it may be easy for us to imagine a future where fleets of autonomous aircraft ferry passengers through the sky. After all, autonomous cars are already starting to take to the roads. But aircraft are a little different: much of the flight is already automated. But it is in exceptional circumstances (the outliers), where the autopilot may not be able to make sense of the situation (possibly because a sensor is malfunctioning), that it pays to have an experienced pilot in the cockpit. This was seen clearly on January 15, 2009, when US Airways Flight 1549 suffered multiple bird strikes shortly after takeoff from LaGuardia Airport. Both engines were disabled. Capt. Chesley Sullenberger made the decision that gliding to a nearby airfield was not an option, but was able to perform an emergency water landing on the Hudson River. Not a single life was lost and the event became known as the ‘Miracle on the Hudson’.

Solving the pilot shortage isn’t a trivial problem. Global training capacity would need to increase dramatically, and systemic barriers to becoming a pilot need to be lowered. The most notable barriers are the cost and duration of training, which are partially baked into regulations. Aspiring pilots must meet a high minimum threshold of flying hours before being deemed ready to fly commercial aircraft. Reaching this threshold, in addition to spending a fortune on tuition at a flight school, is dissuading many of from a career in aviation. But how do you lower these barriers without simultaneously compromising flight safety.

At Paladin:Paradigm Knowledge Solutions, a Canadian aerospace analytics startup, we are taking a data approach to this problem. If data mining techniques can be used to measure pilot competency, then we no longer need to use a time-based proxies as a measure of competency. Years of experience are usually an indication of competency, but not a guarantee.

In order to do data-driven competency-based training, we are looking at unsupervised learning models for ways to compare different simulated flights, performed by pilots of differing skill levels, to each other. For a given maneuver, assuming that there aren’t a thousand equally-valid ways of flying that maneuver, we can establish a small set of “expert” reference flights. These should show commonalities across the many dimensions of flight data. We then have a novice fly the same maneuver and examine the similarity of their performance to that of the expert reference set. There are various techniques for measuring the similarity of two data sets, and these can be used as a measure of competency.

Our approach is similar to that of taking 100 recordings of concert pianists playing the same piece of music and teaching a machine to know what a concert-caliber performance should sound like. Each virtuoso may have their own interpretation of that piece, but with enough expert “examples”, the algorithm can get a sense of what an expert sounds like, despite variations in style. When the novice plays the same piece, the machine recognizes how close to concert-level she is, and can perhaps even suggest a path to improvement.

A flight training curriculum corresponds to many scenarios being flown in the simulator, each testing different skills. While earning their wings at a flight school, a trainee’s learning is split between the classroom, the computer-based training station, various flight simulators, and in the cockpit of various small aircraft. Combining data across these different learning modalities, along with assessments and instructor evaluations, we can begin to measure how each trainee learns best—where they’re excelling, and where they’re falling behind. This can lead to better sequencing and curriculum design, optimizing the time it takes for each trainee to reach proficiency.

The technology required to achieve personalized, adaptive education is still evolving. We’re in the aerospace sector because the need for data-driven adaptive training is clear, and the data being produced is abundant. Other industries are implementing their own version of adaptive training. Knewton is an example of a technology startup seeking to disrupt higher education with their adaptive, online learning platform. Discussions around evidence-based training are happening everywhere complex skills need to be taught in a short period of time, such as in medical schools.

Adaptive learning is an exciting trend. By combining educational data mining techniques and machine learning, artificial intelligence and human intelligence can become forces that shape and improve each other, pushing us to be better at our jobs and helping us attain expertise quicker than ever before, whether that’s flying an airplane or designing the next great app.

Published in
0 Comments

Leave a reply

Log in with your credentials

or    

Forgot your details?

Create Account