March-April_2022 - 24

COVER STORY
The Data-Backed Possibilities of Predictive Maintenance
While data collection and algorithmic technology continues
to advance and move preventative maintenance into the
future, the question is now shifting to, how do you put it to use?
What data is good data? And can you ever collect too much
data?
Joel Klooster, who is the VP of Product Management, GE
Digital, said the answer to the latter question is no - having
a long, robust history of data is never a negative. However, if
you're training an algorithm to change some operations over
to, you won't want to feed it all 15 years' worth.
" The simple answer is there's no such thing as too long to
be storing the data. Now, to be using the data, you want to
make sure that your operation and the type of environment
you're changing in is relevant to the data set that you're using
to train the algorithm. You may have 15 years' worth of data.
You may only want to go use the last couple of years of data
to be training the algorithms because you've got different
operating procedures, different fuel loading procedures,
different weather patterns, city pairs, things like that, "
Klooster said.
Where data collection and analysis help in preventative
maintenance is reinforcing what you know, then discovering
what you know you don't know and what is totally unknown to
you; transcending preventative maintenance into predictive
maintenance.
" There's the concept of known knowns. We know what we
know. Where predictive maintenance really starts to help is the
known unknowns. What are you trying to predict that you don't
know today? Then there's the unknown unknowns. We know
there's things we don't know, but we don't know what they are.
Those things will come up. Having a rich set of historical data
to go back and look at when those things do come up is always
going to be valuable, " he said.
Good Data Makes Good Algorithms
This why feeding an algorithm the correct data is crucial.
Klooster said that they have found that unlike the social algorithms
most people are familiar with, those utilized by the
likes of Google and Facebook which take in mass amounts of
unstructured data, algorithms assisting with aircraft maintenance
need more selective approaches.
" You're looking for very rare events that happen. The risk
of too complex an algorithm is that you end up overfitting to
some random occurrence that happened to happen at the
same time as this failure mode. You end up correlating the failure
to something that has no actual tie to it. You end up overfitting
to something that's not related.
" Being able to have context in the data, understand which
systems are actually interconnected on the aircraft, and have
some understanding of how the aircraft operates, and how the
different systems are built, is pretty important so that you're
making sure you're not finding random correlations in the data,
and end up triggering a bunch of false alerts, which would be
more distracting than not, " Klooster explained.
Clean data is the key. Klooster said they put a lot of effort
into making sure that the data coming off of the aircraft is
filtered of any errant parameters - things that are just outside
the norms of normal boundaries of operating the aircraft.
" There's noise in the data that's captured, and making
sure that you're cleaning it, and finding the right information,
and then building the right correlations between which
maintenance activities were to fix the right failures that have
happened in the past so that you're training the algorithms in
the right way, " he said.
When starting with a software like this, Klooster said there's
no one-size-fits-all approach to start integrating data collection
and algorithms. It's very dependent on the sophistication
level of an airline, but in his mind, it all comes back to being
ready from a change management perspective.
" We will often start working with an airline to look at historiGE
DIGITAL
cal data, and show what could have been, what the algorithms
could have caught, and just make sure that there's trust and
belief in the system, on what it would do. Often, operators will
run these algorithms in the background. Then let things run to
failure to see how accurate those algorithms were, and would
they have caught it? To kind of build that trust and build the
faith in the system, which isn't a bad thing to do just to make
sure that there is that trust and that understanding. Start with
something that's going to be impactful. Start with one system
24 MARCH/APRIL 2022
AIRCRAFT MAINTENANCE TECHNOLOGY
March-April_2022

Table of Contents for the Digital Edition of March-April_2022
