AI in the Sky: How Artificial Intelligence and Aviation Are Working Together

Artificial intelligence (AI – also related to Machine Learning, or “ML” as it’s called) has reached new heights: a cruising altitude of 10,000 - 70,000 feet to be precise. Commercial airlines and military aviation have already begun embracing AI, using it to streamline routes, cut harmful emissions, improve customer experience, and optimize missions. However, with AI comes a string of questions, technical challenges, and even mixed feelings.

Both the Federal Aviation Administration and the European Union Aviation Safety Agency (EASA) have taken a positive interest in AI. EASA published a report in February 2020 discussing the trustworthiness of AI and how aviation can take a human-centric approach to AI programs. Boeing and Airbus are working on AI both separately and via combined international partnerships. The world’s aerospace safety organization, Society of Aerospace/Automotive Engineers (SAE) is publishing aviation standards and training based upon AI (this author’s company, AFuzion Inc., is the primary training resource for all SAE worldwide training programs). But there are still many questions left unanswered, particularly when it comes to safety. With so much unknown surrounding AI, does it have a place in our safety-critical world?

The airline industry may have some answers.

Defining A.I.

A large hurdle that the FAA and EASA have encountered in discussion around AI is that everyone has a different definition of what AI is. How do you define something that is constantly evolving? To start, AI is much more complex than the typical algorithm or program you might use on a day-to-day basis. AI allows machines to learn from experience and adjust the way they respond based on the new data they collect. Traditional aviation software is certified to be Deterministic via guidelines such as DO-178C (avionics software) and DO-254 (Avionics Hardware). But AI essentially enables the same software inputs to yield a different outcome as the software “learns” over time; how can mandatory certification determinism be achieved with a decidedly evolving program to ensure safety?

For example, AI may have helped to develop the algorithms that provide you with customized daily news, or provided you with personalized shopping recommendations based on your search and browsing history. But now we’re talking about AI mapping out your airplane’s flight path—or flying the plane on its own or enabling swarming UAV’s in tight formation to perform a mission. Those tasks are much more challenging for many people to trust, in particular the government and consumers.

EASA’s broad-base definition of AI is “any technology that appears to emulate the performance of a human.” The human-like aspect of AI is frequently included in definitions of AI, and is one of the reasons there have been some questions about the safety of AI. There is always room for human error, so if AI is performing and evolving as a human would, doesn’t that mean there’s room for AI error or safety violations also?

The short answer is: “not necessarily”. Thankfully, AI doesn’t work the way that humans do. Engineers have developed many solutions for deterministic AI learning then monitoring what AI is doing in real time. Many of the real safety concerns come from the cybersecurity sphere, but there remains the challenge of clearly communicating to passengers, pilots, and regulators how AI actually operates. And that’s exactly what EASA and certification authorities/experts are trying to accomplish. EASA has said that a key priority for them is to stimulate international discussions and initiatives—in particular, to coordinate proposals addressing the complex safety and cybersecurity challenges involved in AI-assisted aviation. To accomplish this, EASA and industry are increasing their investment in AI research and technology, while encouraging other countries and entities to follow their footprint in incorporating AI into their aviation industries. This is already underway with AI-based flight planning, simulation, and training; this is paving the way for AI to be gradually introduced into the cockpit. AFuzion believe aviation AI will follow automobiles by 8-10 years so look for meaningful cockpit AI solutions in the 2030’s.

Alaskan Airlines and AI

Though AI has been around since the 1950s, only recently has the aviation industry started to employ AI to streamline and improve airplane performance. The increased interest in AI is mostly due to the demand for air travel. According to the International Air Transport Association, air travel is projected to double over the next 20 years and airlines need to find new ways to keep up with the increasing number of passengers. AI programs could aid with air traffic, managing queues, and enhancing the in-flight experience.

One great example of an airline employing AI is Alaskan Airlines. The company used the slow-down of the pandemic to test out some new flight-path programming for their aircraft. During a six-month trial period, Alaskan Airlines implemented an AI-driven program called Flyways to discover optimal flight paths by factoring in the original route, current weather conditions, weight of the aircraft, and other factors to determine what the most efficient course would be.

Throughout these flights, the AI program tested all the possible routes, collected data on mileage and fuel use, and used that data to refocus its subsequent efforts, all with the goal of creating the most efficient flight route in real time.

“This is what machines are really good at, taking huge data sets and putting them together,” pilot and head of corporate development at Alaskan Airlines, Pasha Saleh, told ABC News. “Flyways is probably the most exciting thing that I’ve come across in airline technology since I can remember.”

During the six-month pilot program, Flyways shaved an average five minutes from flights. That may not seem like much, but that amounts to a whopping 480 thousand gallons of jet fuel saved, which for Alaskan Airlines, was a major win as the company tries to meet its pledge to be carbon-neutral by 2040.

The DO-178C Safety Net

The main concern around implementing AI into transportation services is safety. Many entities, including the FAA and Department of Defense, look at AI through a “guilty until proven innocent” lens. One fundamental aspect of safety-critical systems is consistency: explicitly proving that the same inputs provide the same outputs, every time. This is where DO-178C comes into play.

DO-178C is a set of guidelines covering 71 Objectives to ensure that software will perform safely in an airborne environment. The guidelines categorize software on five levels of Reliability, ranging from “No Safety Effect” to “Catastrophic.”

Not only does DO-178C provide safety measures, engineers have been working on some technological solutions to help make AI safer and to keep it in check. Some of these include:

• Installing an external monitor to assess the decisions of the AI engine from a safety perspective
• Building redundancy into the process as a safeguard
• Reverting to a default safe mode when unknown or dangerous conditions occur
• Reverting to a full static program so that AI cannot evolve on its own. Instead, the AI would perform a safety analysis after the program is run and determine whether the program is safe

Along the same vein, EASA has made additional suggestions to ensure AI safety:

• Keeping a human in command or in the loop
• Monitoring AI through an independent AI agent
• Examining the AI output through a traditional backup system, or safety net

The key takeaway here is that there is a great deal more work to do in order to monitor AI and ensure the proper level of safety, but AI represents one of the most promising developments in aviation today. Harnessed properly, AI could help ensure a sustainable future for the aviation industry amid continued rapid technological advances.