Artificial intelligence specialist Daedalean and the European Union Aviation Safety Agency (EASA) have begun the second stage of a joint research project to support plans for the regulator to certify the use of AI technology in autonomous aircraft. The work, which is covered by an Innovation Partnership Contract, will implement and refine the concepts developed in the first phase of the research with a view to publishing guidelines in early 2021.
In January 2020, experts from Switzerland-based Daedalean and EASA completed a 10-month study into the challenges posed by using neural networks for aviation applications. This covered the use of machine learning in aircraft and creating standards for employing this aspect of AI in safety-critical avionics.
The research informed a joint report called "Concepts of Design Assurance for Neural Networks" (CoDANN) that was published in March 2020. This built on EASA’s AI Roadmap, published in February 2020, which itself drew on the wider 2019 European Commission document called "Ethics Guidelines for Trustworthy AI."
The CoDANN report included an outline of realistic performance and safety assessments to define the failure tolerances and datasets needed to support appropriate safety levels for AI technology. According to Daedalean, the quantitative analysis conducted so far has demonstrated that it is feasible to guarantee neural network safety at the levels of criticality required in aviation.
The next report, based on the second phase of the joint research project, will propose high-level guidelines for neural network-based systems and how these can be included in safety assessments and explained to industry stakeholders at a practical level.
“We will be taking a concrete inflight system through a certification trajectory to find the open questions, with the intent to provide concrete usable answers,” said Daedalean cofounder and CEO Luuk van Dijk.
The company said it will provide more information at a later date about the "inflight system" being used for the research and on other use cases being evaluated.
