AOPA Pilot Magazine - August 2021 - 99

frequency of about one cycle per 10 or
11 seconds, six times faster than the one
cycle per minute oscillation that FEVA
1.0 was designed to look for. In fact,
FEVA 1.0 was specifically programmed
to ignore EGT oscillations faster than
one cycle per 30 seconds based on the
" expert assumption " that oscillations
faster than that almost had to be noise.
No wonder it didn't alert on Wally's failing
exhaust valve.
So much for the expert. (The expert
was me. Sigh.)
Would the machine learning model
have done better? We wondered the same
thing, so we ran Wally's flights prior to
the 2018 annual inspection through
FEVA 2.1's machine learning model. The
results are at left.
Wow! The machine learning model
returned an off-the-charts risk score for cylinder 2, one of the
highest risk scores we've ever seen the model produce. How
exactly did it conclude that the number 2 cylinder was at extraordinarily
high risk for exhaust valve failure? That's hard to say.
The model uses more than 35 input variables to make its preThe
proof is in the pudding
To illustrate how far we've come in the past seven years, a case
study may be instructive. In 2015, Wally purchased a beautiful
2008 DA40 Diamond Star. Wally has been religious about uploading
the data captured by his Garmin G1000 to the SavvyAnalysis
platform. In July 2018, Wally put his airplane in the shop for
its annual inspection. During the compression test on the airplane's
Lycoming O-360, the compression on cylinder number
2 was 0/80. A borescope inspection revealed a seriously burned
exhaust valve in that cylinder that probably would have failed
soon had it not been caught at the annual.
Leading up to the annual, each of Wally's flights had been
scanned by FEVA 1.0, and none raised a red flag that the number
2 exhaust valve might be failing. Why not?
In 2018, FEVA was still an expert system programmed to
alert whenever it found a slow, rhythmic EGT oscillation that
we knew was reliably indicative of a failing exhaust valve. If you
look at the EGT data from Wally's last flight before the annual
inspection (top), you can see that EGT number 2 looks wonky.
But it's not the sort of slow, rhythmic oscillation that FEVA
1.0 was programmed to look for, but rather a much faster rhythmic
oscillation. In fact, EGT number 2 was oscillating at a
dictions. It looks for hidden patterns in the data, some of which
might not be apparent to the human eye. While the machine
learning approach we use is enormously powerful, one drawback
is that it is difficult to tease out the effects of individual
variables to arrive at an explanation of why the model made the
prediction that it did.
What we can say is that the model did not have access to
information from the compression test or the borescope inspection.
It looked solely at engine monitor data, and it looked over a
series of flights in order to come up with its prediction that the
number 2 exhaust valve was at high risk of failing.
This is exciting technology, and I think we've just scratched the
surface of what it can do for piston GA maintenance. My team is
working hard to make predictive analytics smarter and more accurate,
and to extend it to other safety-critical areas such as sticking
valves, failing ignition, and detonation/preignition. If you're flying
a piston aircraft, predictive analytics might save your bacon
someday. mike.busch@savvyaviation.com savvyaviation.com
Ask the A&Ps /
Every month, experts Mike Busch, Paul
New, and Colleen Sterling answer your
toughest aviation maintenance questions
on our Ask the A&Ps podcast. Submit
your questions to podcasts@aopa.org.
aopa.org/news-and-media/
podcasts/podcasts/ask-the-a-and-ps
AOPA PILOT / August 2021
99
http://www.savvyaviation.com http://www.aopa.org/news-and-media/podcasts/podcasts/ask-the-a-and-ps http://www.aopa.org/news-and-media/podcasts/podcasts/ask-the-a-and-ps
AOPA Pilot Magazine - August 2021

Table of Contents for the Digital Edition of AOPA Pilot Magazine - August 2021
