Avionics News February 2023 - 36

PIONEERING ARTIFICIAL INTELLIGENCE AVIONICS
Continued from page 35
EASA and Daedalean formulated a modification to the usual
V-shaped software assurance process used in avionics (and in
other safety-critical fields) to ensure the AI computer learning
meets standards. The traditional V specifies a systematic set of
activities that ensures software processes and products conform
to requirements, standards and procedures. It applies to the
life cycle of developing a specific application - the stage of
writing the code itself. It includes verification of the product
requirements and validation of code as fitting the requirements
(see illustration below).
EASA's AI task force team and Daedalean proposed a
W-shaped process for AI and machine learning avionics with
the middle of the W being used for learning process verification
(see illustration on opposite page). The W ensures the machine
learning technique is applied correctly and offers proof that the
resulting application will work as intended and give the correct
answer in the number of cases not less than required by the
regulator. For example, 99.xxxx % of cases with the exact limit
set depending on the safety criticality of the function involved.
A second EASA/Daedalean 136-page report with the same
title followed by the designation II came out in 2021. The
second project focused on further steps needed to pave the way
for the certification of the first applications of AI and machine
learning in safety-critical avionics. EASA issued its guidance
for machine learning applications describing the objectives
of compliance in December 2021 and what is involved for
learning assurance. The guidance describes the W and requires
compliance with it.
Daedalean also participated in a joint research project with
the FAA William J. Hughes Technical Center Aviation Research
Division on the use of neural networks in general aviation.
This produced a 140-page report titled " Neural Network Based
Runway Landing Guidance for General Aviation Autoland "
(see DOT/FAA/TC-21/48). The study focused on the machine
learning visual landing system for fixed-wing aircraft developed
by Daedalean to see if it could serve as a backup for other
navigation systems such as GPS during an outage. The flights
used a forward-facing camera on an experimental Cessna 182.
(For a profile of the FAA Tech Center's history and current
research activities, see " FAA Tech Center: Humble Origins to
Cutting Edge " in Air Traffic Technology International 2023 by
David Hughes.)
Design assurance process
36 avionics news * february 2023
" During the test flights, we flew standard landings and
robustness tests of our systems, " said Luuk van Dijk, chief
executive officer and co-founder of Daedalean. " This
included two flights and 18 approaches, both over previously
trained and untrained runways, in various conditions. The
flights occurred in degraded daylight with steep glideslope
angles and sharp roll maneuvers. The system performed
well; in particular, it spotted a runway that was not part of the
training set from almost 3 nautical miles away. "
Van Dijk is a software engineer with a doctorate in physics
who worked as a software engineer for Google and a flight
software engineer for Space X. Anna Chernova is the cofounder
and chief product officer for Daedalean and is a
private pilot. She was a quantitative analyst at Google.
At Daedalean in Zurich and in Latvia, the company has
conducted a series of data collection flights and simulations
to gather a training dataset of images for use in machine
learning. All told, the company has gathered 1.2 million
images from 500 hours of real flight video records.
This included 7,000 encounters between a data
collection aircraft doing the recordings and another flying
the encounters as the target. Daedalean rents aircraft and
helicopters for these missions and works with helicopters
used by its partner Embraer in Rio de Janeiro and a Cessna at
Avidyne. It recently purchased a Piper in Switzerland.
During 10 to 20 seconds of tracking the other aircraft, the
camera makes six images per second. The recording and
target aircraft play different scenarios
approaching each other from different
directions and at different altitudes,
directions and speeds. The company
also records ADS-B In and GPS
data for comparison purposes when
evaluating its algorithms. Post-flight,
there are several algorithms and
humans analyzing the video second by
second and image by image.
Specially trained data analyzers
identify aircraft in the image. Then
a huge statistical analyzer called a
neural network analyzes each pixel
and relates these pixels to each other
to answer the question, " is there an
Avionics News February 2023

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