What was once a dream is now becoming a reality. Where some see only financial savings, others see an almost insurmountable challenge for airlines. Will AI be able to achieve the same feats as Chesley Sullenberger? The pilot who successfully ditched an A320 on the Hudson River. 

 

AI on board, a feasible concept ?

According to a UBS study published in 2017, autonomous aircraft could arrive on the market as early as 2025 and would save airlines as much as $35 billion. Artificial intelligence will increasingly assist pilots during flight. Provided that we are able to certify systems with autonomous decision-making.

Over the years, the growing maturity of artificial intelligence has called into question the central role of the pilot, who is still the sole master of the aircraft. His vocation today is to ensure the safety of the passengers and to fly the aircraft in the best possible way in all circumstances. It already shares its work with the computer, since most civil flights today are on automatic pilot. But artificial intelligence promises to go one step further, changing the way they operate.

As the investigation into the crash of the Rio-Paris A330 in 2009 states: “The accident took place in cruise flight at high altitude, at the crossroads of the intertropical front. As the aircraft passed through a severe thunderstorm, the pitot probes became iced up and the airspeed indications were momentarily lost. Inappropriate pilot reactions caused the aircraft to stall until impact. The role of AI would be, in the near future, to apply the right process to each technical problem envisaged.

 

Processing massive amounts of data

Every year, several zettabytes of data are produced, stored and exploited. This means billions of billions of bytes, a wealth of information that allows the emergence of new tools. These tools use disciplines such as machine learning or deep learning, combining computer science and mathematics to design algorithms for processing massive data, which have many industrial applications.

Since this data is too voluminous for a human to be able to handle every stage of processing, and because we also want to see the emergence of innovative services and lessons, without any preconceived ideas, the idea has arisen to allow machines to learn automatically. The challenge of machine learning, fed by probabilistic modeling and optimisation and served by a theory of learning, is therefore to provide guarantees on the robustness of the results. And its automation and prediction capabilities offer great potential benefits for aviation.

 

Overcoming human error

Indeed, learning algorithms alone select the optimal model to describe a phenomenon from a mass of data with a complexity unattainable for our human brains: up to several million variables. In the context of aeronautics, this means that the AI at the controls of an airliner will be able to determine the right process to apply according to an engine failure, the deterioration of pitot probes, a fire in the cockpit, an alarm that sounds, an area of excessive turbulence or an event that requires an emergency landing. But human error will no longer be taken into account, often caused by contextual factors. Indeed, the machine does not need to sleep, feed, or have any alterations to its ‘consciousness’. Depending on the variables it is given, it applies an automatic response and that’s it!

Based on the example of the Rio-Paris crash, and according to the BEA report (Bureau d’enquêtes et d’analyses pour la sécurité de l’aviation civile) and journalistic investigations, it turns out that the 3 pilots on board the A330 were not in optimal condition to ensure the flight. Already on take-off, the pilots were showing signs of exhaustion: “Cockpit recording: captain at 1 h 4 min 19 s” Last night I didn’t sleep enough. One hour was not enough before” and “during the first 23 minutes of the recording, silence dominates the crew with radio communications from Recife control in the background, attention is relaxed to the point of listening to music”.

There is nothing to suggest that the artificial intelligence would have been able to prevent the aircraft from stalling. But it would not have shown “signs of exhaustion”, because it is… a machine! And the difference with autopilot is that artificial intelligence is able, in the event of a breakdown, for example, to decide on a change of control mode and to change actuators almost in real time.

 

Technological and regulatory challenges

However, there are still a number of questions to be answered before we can claim victory, and negotiations with pilots’ unions and regulatory bodies are also to be expected.

Moreover, the automation of aircraft offers many advantages: faster decision-making, processing and scheduling of mass data, management of complex and coordinated manoeuvres, etc. All of these benefits offset another challenge, that of regulating the airspace, which is currently saturated. As promoted today by the European SESAR project, the objective is to provide Europe with modern systems, in order to offer pilots real-time information to adapt their decisions, thanks to a network capable of ensuring the optimised regulation of the whole.

The major challenge of unmanned aircraft is, first and foremost, to ensure the safety of the aircraft and the reliability of the AI-based systems. In an aircraft, there must be less than one catastrophic failure per billion flight hours, an international standard! One of the possibilities lies in hardware and software redundancy.

But beware of addiction to automatic systems, which can lead to complacency, in other words, the risk of over-reliance on machines. Therefore, everyone must find their place on board! It remains to be seen whether we, as passengers, have more confidence in the man or in the machine…