Air Traffic Management (ATM) provides a highly safe framework for aircraft to fly thanks to skilled controllers who operate robust systems. However, ATM can still improve in terms of operational efficiency while maintaining its current safety levels, as Eurocontrol reported in the last ATM cost-effectiveness Benchmarking report of 2018. Machine learning technologies can analyse vast amount of diverse data (usually from different sources) to extract hidden patterns and information. This knowledge can then be used in predictive capacities (for instance, on safety events, training or performance deficiencies) that then support operators’ decision-making. An air traffic controller, as any professional, learns and further strengthen his or her skills through practice. Artificial Intelligence can process years of this data, resembling this idea of practice, and put this knowledge at the service of the human operators, increasing efficiency in the operational room. Following the coronavirus pandemic, Air Navigation Service Providers (ANSPs) will face some years of cost-control, perhaps lower staff and, therefore, challenges to their operations; they may even be requested to reduce ATC fees. This operational and financial burden, along with lower levels of traffic, may pave the way for new AI systems and AI tools that allow ANSPs to scale back capacity with lower costs and increased safety. The SafeOPS project, coordinated by TU Munich in partnership with Innaxis, DFS, Iberia, Pegasus and DeepBlue, addresses this objective by focusing its research on the “from prediction to decision” paradigm by expanding the current ATM system with an information automation-based decision intelligence.
Safety is the priority of any proposed improvement in aviation. Even with today’s incredibly high safety levels, human operators learn how to manage uncertainty. Currently, air traffic controllers make decisions while assuming the information they receive is accurate. With such information, they accept a degree of uncertainty, such as weather or aircraft intent. This uncertainty, in today’s ATM systems managed by ATCOs, is not mathematically quantifiable as a probability distribution. SafeOPS will develop an integrated model of risk for the selected safety-critical scenario: Go-Around. The computed risk is added information that flows into the planning and operational management of the overall ATM system.
On the one hand, increasing levels of digitalization lead to more data sharing among stakeholders and technology providers, which in turn requires a stronger degree of vigilance towards potential cybersecurity threats. Secure data platforms, suited with encryption technologies and powerful tools to monitor data flows and the performance of the system, are required to prevent potential cyber attacks.
Lastly, there are risks associated with the criticality and complexity of the controller position being represented in software. Most current systems have millions of lines of code and AI will be an invaluable tool for developing feasible roadmaps that adapt these systems to higher levels of automation.
In spite of the risks, the advantages of applying AI techniques in support of current aviation technologies and procedures are clear. While progress on these applications is demonstrating its added value in assisting current performance, there are still roadblocks to implementation for different aviation stakeholders.
-ANSPs: Current ATC systems have long innovation cycles and upgrades require long and expensive processes. AI technologies require more agility and shorter, simpler cycles for innovation. In return, ANSPs leading AI technology will become more cost-effective and will gain the experience in automation needed to address more complex challenges and beat capacity limitations when pre-pandemic levels of traffic come back.
-Regulators and certification: The aviation industry is already pushing for the first AI-based system approved for operations in less than a decade. The demand on manufacturers to offer different ML solutions that enable more cost-efficient flights is more relevant than ever in this post-COVID era. EASA has already published a first roadmap for AI solutions including tentative certification milestones.
-Air Traffic Controllers: For end users, it is very important to understand how new technology fits into existing processes and to make them part of the solution in case those processes need to be changed. Specific training will be needed to interact with AI-based predictive technologies, to incorporate its associated uncertainty into operations and to better present information. Users are a key, active element in AI work.
To address these issues, SafeOPS incorporates end users as full partners of its consortium (DFS, Iberia and Pegasus): SafeOPS involves both the ATM and airline operations worlds to identify possible hidden safety risks. Collecting users’ requirements from the earliest stages and having their support to define the operational problem is essential for the project. Collaboration among ML engineers and domain experts facilitates the specific definition of the problem, improves the model performance and facilitates its adoption. This collaboration will support SafeOPS in addressing the human factors linked to the adoption of increased levels of automation. Beyond “information overflow”, the ATM human agents will have to adapt to more information provided by big data analytics, precise but sometimes provided in probabilistic terms. Clever HMI refinements will certainly help mitigate the potential overflow of information. However, research on the impact of information automation on the ATM system needs to be conducted.
Several ANSPs are asking for higher flexibility in technology architectures and AI will be a key piece of the next generation of systems. Modernization programs in ATM will come with tools in which AI will be a crucial component to efficiency and safety improvements in the operational room. SafeOPS aims to demonstrate the feasibility of the concepts from a technical perspective in cooperation with the operational experts. The objective is to make sure the solution developed responds to a real operational problem and matches users’ requirements and needs, paving the way towards a future implementation.
Concretely, the scope is to demonstrate that a Go-Around can be accurately detected in the data (data labelling) and predicted with enough time for the controller to prepare and react. Our solution will be tested in both a qualitative and quantitative manner. The qualitative part will include workshops and tactical simulations to evaluate the trust of ATCOs in the concept. For the quantitative part, a comparable framework will be developed to ensure that the resilience and safety of the system are not jeopardized by the predictive capabilities. In any case, additional research, development and validation work will be required before implementation and commercialization of the solution. The SafeOPS consortium is committed to fully aligning on initial research to enable its future adoption; in parallel, significant progress in certification of AI technologies would be required from authorities.
