Perfecting human-machine collaboration

Artificial intelligence is the focus of the first whitepaper in the series.

“The workgroup looked at what technologies would have an impact on air traffic management – and AI is clearly one of those,” said Patrick Souchu, DSNA Director of SESAR Programmes and Chair of the CANSO Strategic Technology Workgroup that’s overseeing the whitepaper series. “Applications based on AI are coming to the fore, but we don’t know yet how far the technology will develop at this stage.”

AI is a broad term that has evolved over time since the technology first came on the scene in the 1950s. It is the recent advent of a “tsunami of data” – as Béatrice Pesquet, Research and Innovation Director at Thales Air Mobility Solutions and lead author of the whitepaper puts it – that has accelerated development and made the technology so important to future air traffic management operations.

AI is usually defined as any technology that mimics the performance of a human. Key to this is machine learning (ML). The important point to note about ML is that its algorithms are determined by the data. As such, it is very different from traditional software programming based on rules and pre-conceived models.

As the data keeps coming in, the algorithms change. The machine learns. And though this is a massive technological leap, it also carries a host of challenges for its use in the air traffic management (ATM) environment.

For example, the “learning” must happen offline at the moment. Aviation is based on strict safety protocols and the output of any system must be validated. Performance is constantly checked, and any machine learning algorithm must be verified.

Pesquet identifies three elements that will be important for ML-based technologies to gain the trust of air navigation service providers (ANSPs) and air traffic controllers (ATCOs).

Most crucial will be the certification of these technologies. EUROCAE (WG114) and SAE (G34) are guiding the relevant authorities to agree industry standards on aeronautical systems using AI. These are now in place making possible certification that the industry can trust.

AI-based systems will also need complete and precise design. Data is part of the algorithm specifications and so such factors as data accuracy, completeness and relevance are extraordinarily important to overall performance. The technicians involved in the design need to not only understand the complex engineering involved but also the business for which it is intended.

“This is something very new,” says Pesquet. “Air traffic controllers need the right information at the right time.”

A third element is “explainability”. Essentially, ATCOs will want to know why an algorithm has provided a particular answer and how that answer compares with the manual completion of a task. This will require ongoing training and monitoring.

Three steps to trust

Souchu stresses that the ultimate aim is not to replace human controllers but rather to optimise human-machine collaboration.

“This is not like a driverless car,” he says. “AI won’t take the human out of the loop. What it will do is alter responsibilities and give humans more time to concentrate on critical issues. AI is more about freeing humans from repetitive tasks where mistakes can be made. This increases the resilience of the system and overall safety.”

The whitepaper further explains: “A human-machine collaboration can also be deployed as a feedback loop to continuously improve the AI system with more incoming data, which would be particularly useful for domains in which the operational dynamics change through time.”

The volume of traffic usually makes it resource intensive to evaluate all flights, for example. A ML system can, however, do a complete analysis of all flights and pre-classify any anomalous flights. This pre-classification can be studied by human experts and additional information added, such as causal and contributory factors that will facilitate decision making.

Human-machine collaboration

AI use cases are in progress in non-safety critical scenarios. The aim is to provide better information in a more precise way and so improve situational awareness. Weather impact is an obvious example.

Work will eventually target the automation of certain processes and even real-time involvement in safety critical moments. Pesquet says that the latter should happen around 2025.

Access to data will be critical to the timely introduction of AI. And Souchu admits this is easier said than done. “Data sharing is not always easy,” he says. “Airlines are sensitive about data sharing, and it is important that the industry deals with confidentiality issues. Remember, airports and airlines are using more AI too so collaboration and data sharing will be vital.”

Training and attracting talent will be equally crucial to AI success. Industry partners stand ready to help with training and, as Pesquet notes, “acceptance is related to understanding so ATCO training will pave the way for the full benefits of AI technologies”.

She is also confident that ATM is an attractive industry for technology students. “It is a user-centric domain with enormous potential for change,” she concludes. “The younger generation want to see a greener, efficient world and they can play their part by integrating AI in air navigation services.”

Emerging technologies for future skies: Artificial Intelligence is available via the CANSO publications library.

The next steps

New and emerging technologies are essential in meeting the needs of our future skies. They will not only deliver operational efficiencies and safety enhancements but also help accommodate increasingly diverse airspace users and boost the sustainability of aviation.

But the industry must learn to harness innovation if it is to reap the potential of these technologies.

A group of aviation industry thought leaders and technology experts has come together under the guidance of CANSO to examine new and emerging technologies. Together, they are sharing their thinking through an exclusive series of whitepapers on breakthrough technologies and the role they will play in the transformation of our industry.

The whitepapers will cut through the hype around new technologies, and showcase what they are, what they do and the role they should play in aviation’s future.

The key topics being examined are artificial intelligence (AI), system wide information management (SWIM), space-based CNS, blockchain, airborne capabilities, virtualisation, and new concepts stemming from UTM. 

skyguide continues to evaluate artificial intelligence. The immediate aim is to share the same situational awareness with air traffic controllers in En-Route operations through the AI Situational Awareness Foundation for Advancing Automation (AISA) research project.

The project envisions an AI-driven system instead of isolated tools such as conflict detection or coordination systems and will look to optimise human-machine collaboration.

In an upcoming AISA experiment planned from November 2021 to January 2022, skyguide will explore such critical questions as:

• Is AI situational awareness comparable to human (ATCO) situational awareness?
• How does AISA react to new situations?

HungaroControl is using AI to improve simulation training and operations.

Virtual Pseudo Pilot is a software solution for air traffic control simulators that performs pilot tasks using advanced voice recognition technology. In essence, the software recognises ATCOs’ voice commands, acknowledges those commands in the correct manner and then executes the instruction in real-time.

In terms of operations, the AI-based DeFOG tool is compatible with any video surveillance system. The tool enables controllers to guide aircraft even in the foggiest conditions.

HungaroControl also reports that AI will “improve dynamic airspace configuration and support decisions on when to open or close sectors. AI-powered algorithms can show the minimum number of sectors that need to be kept open and so optimise the workload of air traffic controllers”.

The ANSP also sees a role for AI in the introduction of unmanned traffic management.

Air Traffic Management Research Institute (ATMRI) at Nanyang Technological University, Singapore, is pioneering a hybrid AI-Human ATM approach. In hybrid AI-Human ATM, machines will be able to co-learn safety-critical tasks from humans and evolve into reliable AI-based assistants for air traffic control officers (ATCOs).

ATMRI has developed several solutions, including airspace capacity prediction, airway network optimisation, and the identification of unstable configurations in terminal airspace.

For example, ATMRI has developed AI algorithms for mitigating delays in Terminal Manoeuvre Airspace (TMA). Results show that, by applying minor speed adjustments starting at 300 nautical miles (NM) from the airport, up to 85% of predicted delays were absorbed in the cruise phase. As a result, the average delay per flight decreased from six minutes to less than two minutes.

ATMRI has also built a Deep Learning model that performs aircraft trajectory prediction by incorporating short-term aircraft tactical intent and long-term historic flight profile data. The proposed model does not require explicit information about aircraft performance and wind data. The model improves the prediction performance up to 30%. It also guarantees the mean error growth rate with increasing look-ahead time to be lower than 0.2 nautical miles per minute.

DSNA is studying AI for optimal sectorisation. In the dynamic airspace management domain, ML/AI techniques can introduce a significant improvement. The basic idea is to train a machine with traffic data and sector configuration and let it learn how to re-shape sectors boundaries to better accommodate the traffic.

In the context of improved performance for flow management at ACC level, the SINAPS project (SWIM Integrated Network management and extended ATC Planning Services) developed in SESAR has delivered a solution for DSNA which was deployed in 2019 in an operational environment in France. The SINAPS project supports both tactical and strategic operations. The next step will address cross ACC optimisation.