Avionics News February 2023 - 35

STORY BY DAVE HUGHES
ence in designing installation-ready avionics hardware that can
pass FAA and EASA certification. AI is new to these regulators,
so certification presents some unique challenges for the partners
to deal with together.
Daedalean has made its own visual AI software and hardware
that has been validated in flight and during simulation and then
demonstrated in real flights to a variety of customers for evaluation
and possible integration into aircraft and products.
The pilot-aid technology can visually detect and assess
in-flight hazards and integrates with TCAS, TAS and other
traffic sources into a single, intuitive human/machine interface.
Daedalean expects a hundred installations a year of its system
once it is certified and available under supplemental type certificates
covering a selection of aircraft models.
The Daedalean developmental system includes one to four
cameras and a computing box. A one-camera system weighs
13 pounds, and a three-camera
system weighs 17.6 pounds.
The system meets DO-178C for
software and D0-254/DO-178C
DAL C for hardware. DAL
stands for design assurance level and defines the amount of rigor
required in development and testing. Level C is major, which
means failure of the system could reduce the safety margin and
increase pilot workload. Daedalean's developmental system is
a Level 1 AI/Machine Learning system as defined by EASA.
Level 1 AI/ML is for human assistance, Level 2 is for human/
machine collaboration, and Level 3 is for a more autonomous
machine. Daedalean said DAL C is appropriate for its Level 1
system.
Avidyne and Daedalean's joint project is Level 1, while Level
3 systems are not even close to actual certification, according to
the Swiss company. Daedalean plans to have a Level 3 system
by 2028 for electric vertical takeoff and landing aircraft, and the
technology will be aircraft agnostic.
The eVTOL startups have an even bigger challenge than
certifying machine learning-based avionics. They also have to
certify new aircraft, flight controls, noise, engine/fan reliability
and other things. It is possible that, for general aviation, the
applications for advanced pilot assistance could come to market
sooner than AI on eVTOLs, according to Daedalean. Avidyne
is the prime example. Daedalean adds that Level 2 (human/
machine collaboration) is not dramatically different from Level
1, except for the interface. The core technology stays the same,
and the certification upgrade may not likely be that difficult for
piston aircraft. Jets are a different story because they operate at
different altitudes and speeds, mostly on instruments.
Daedalean worked jointly with EASA on the certification
concepts to be used with AI and machine learning avionics,
and the organizations published two papers together. This
includes " Concepts of Design Assurance for Neural Networks, "
a 77-page public extract from a 110-page report completed in
2020. The extract outlines safety guidelines for the use of neural
networks in safety-critical avionics. (A Google search of this
title links to this EASA report). The EASA announcement on
the study said Daedalean works with eVTOL companies and
aerospace manufacturers to specify, build, test and certify fully
autonomous autopilot systems and has demonstrated " crucial
early capabilities on a path to certification for airworthiness. "
Several eVTOL companies are flight testing Daedalean technology
to see if it fits their needs.
Some general aviation aircraft
manufacturers are doing the
same. eVTOL makers include
Embraer's Eve Urban Air Mobility,
an independent company developing an urban air mobility
ecosystem. Eve is flight testing scale models and performing
simulations of an air vehicle that will have eight electric rotors
for vertical propulsion and two for horizontal propulsion. Eve
wants to market this air vehicle in 2026.
The difference between neural network software and classical
rule-based computer programs is that the neural network
software isn't written by a software engineer command by command.
The process instead involves a special algorithm tuning a
complicated statistical model by analyzing millions of images.
This is the machine learning process.
Where a human notes there is an aircraft in an image, a statistical
model extracts patterns and finds millions of parameters
distinguishing the images with aircraft using machine learning.
The computer vision looks at all the megapixels all the time,
which gives it an advantage over humans who have to turn
their attention to a particular spot to see the most detail. The
machine-learning algorithm learns to detect patterns of pixels
that human annotators have labeled " this is an aircraft " in the
training datasets. It uses the same information the human sees
on the screen when presented with the camera image.
Continued on following page
avionics news * february 2023 35
Avionics News February 2023

Table of Contents for the Digital Edition of Avionics News February 2023
