Subscribe
Share
Share
Embed
"Smart Metals Podcast," hosted by Luke van Enkhuizen and Denis Gontcharov, offers a clear and practical look into the metals industry's journey through digital transformation, Industry 4.0, and the integration of the Unified Namespace. Listeners can expect in-depth discussions that break down these complex topics into understandable segments, actionable insights, and real-world applications. Luke and Denis bring their expertise to the table, guiding you through the evolving digital landscape with advice on leveraging technology for streamlined operations. Each episode aims to empower metal industry professionals with the knowledge needed to confidently adopt digital innovations and understand the impact of the Unified Namespace in creating a more connected and efficient production environment. Join us to navigate the future of the metals industry with clarity and confidence.
Luke van Enkhuizen: welcome back
to the smart metals podcast, the
show where we talk about digital
transformation for the metals industry.
My name is Luke van Enkhuizen and
I'm together together here with.
Denis Gontcharov: Denis Gontcharov
Luke van Enkhuizen: Today, we will be
talking about the future of your factory.
Particularly, we'll be talking about
something a little bit controversial,
a bit of a thought experiment, where we
dive into the value of domain knowledge
and how important that is for the future
of your factory and how this will maybe
change if we will use a more unified
architecture, such as unified namespace.
So let's kick it off there and
let's talk about this today.
Denis Gontcharov: Yes.
I'm personally very
excited about this topic.
We are entering an age of AI and I think
it's best illustrated with this quote.
It's actually quite old.
The quote reads that the factory
of the future will have only
two employees, a man and a dog.
man will be there to feed the dog
the dog will be there to keep the
man from touching the equipment.
a quote from Warren G.
Bennis.
I think we can't deny that we are
entering a time where more and more
of the thinking and decision making
is being outsourced to machines.
For example, scheduling decisions, process
control has already been automated.
before to make these kind of
complicated process decisions,
you needed to have an engineer who
had sufficient domain knowledge to
operate machines and the process.
given that we are connecting
systems into one unified namespace.
We can imagine having an artificial
intelligence that will take some
of this decision making on it.
What do you think, Luke?
Luke van Enkhuizen: Yeah,
this is actually, I think
already happening, right.
In other ways, in many
other tangent industries.
I mean, like in the financial
services markets, a lot of things
already done by algorithms and rules.
And I think we see these rules
already around us in many
other industries as well.
So I think it's a matter of time until
this really breaks through also in
the metals industry on a larger scale.
The question is only how this happened
and what do we need to get there?
I guess.
So it's not the question of if,
but more like when and how, and
what do we need to get there?
Denis Gontcharov: I really like
your example about the financial
industry, because indeed it's
also a very high stakes game.
If you make a mistake, you lose a lot of
money and financial markets are arguably
more complex because you don't have
physical laws by which nature operates.
So you're dealing with
more stochastic processes.
And if they can't do it, well, why
can't it happen in manufacturing?
Luke van Enkhuizen: yeah, exactly.
And also like it, it's not then
the, that's the systems are using
anything else than we are using.
We are both using data.
We are both using process values instead
of it's the change of price of a stock
or an index or anything like this.
\It's the change of a, the value of.
A temperature, for example, or in a
more discrete sense, it's the time
that a certain operation took and
which parameters were part of that.
So eventually you can plot everything on a
time based data series, and it should give
you some insights for decision making.
Right.
Denis Gontcharov: For sure.
And previously, these decisions
were always made by a human
mind with lots of experience.
I'm thinking of a process engineer
who studied thermodynamics or studied
discrete manufacturing supply chain.
So, but let's pause for a second and try
to define what's in this person's head.
Look, how would you
define domain knowledge?
Luke van Enkhuizen: Yeah.
It's like the know how I think.
In my industry, in the metal
fabrication industry, it's
also called tribal knowledge
sometimes even it's like the, the, the
things that you know among each other
about how things supposed to be, how
the processes are, what means what,
What does it mean that the temperature
has been in this range for so long?
Does it indicate that the machine
needs some changeover soon?
It needs to repair or not?
Does it normal that the
machine stands still?
Does it tick on a funny way?
I don't know.
There's so many ways you can exemplify,
but also what kind of customers need.
What kind of things
customers need from you?
There's so many things that you know,
your company as your domain, that's the
domain you're in and the knowledge that
that sets the rules of the game that you
internally have to find for yourself.
And whether it's that it is in a
document perhaps it's a process
documentation whether it's, it's a
rule book, I think it's a data set.
It could be anything in
between, but in summary, the
things that make your company.
What you are today.
And maybe you can call it
even your secret sauce.
Denis Gontcharov: Yeah, absolutely.
And I think we have to really
make a distinction between
information and experience.
Because at the end of the day, all
the information is technically there.
You have all the sensor data, you
have your rule books, you have your
manuals, but you still need the
person that has the experience.
And I think the way to build domain
knowledge is by blood, sweat and
tears, by studying the information
and applying it in your work.
And that's why it's so hard to
find and so hard to replace.
I think that's why also Boeing
has a lot of issues now that they
are losing the domain knowledge.
By experienced engineers retiring.
Luke van Enkhuizen: Yeah.
It's funny you mentioned blood,
sweat and tears of course, we don't
want to have blood, sweat and tears
in a factory, I hope, but yeah, the
essence is very clear that there is
a history that you've built up about
certain smart ways of doing things
whether it's fabricating parts, whether
it is your continuous process, there
are all kinds Smart things you have done
over the years to make it better, to give
you that edge over the competition or
avoid quality issues, but all of these
things are happening over time and all
these things are in the history of the
company, it says that that may be not
available for the rest of the company
or even the group to do something with.
So AI will be miles away if
we have no way to give that
information further to an AI system.
So you might have lots of experience, but
if there's just no data, if it's siloed
away somewhere, you cannot access it.
And if you're also not recording it,
there's also not much to analyze upon.
And so then it comes to be the question
of what is the physical reality, which
you cannot change, which is something
that is just there, which you can measure.
And which are decisions that are
being made during the course of a
production run, for example, that
influenced that physical reality.
And those are two different things.
Usually they are called actually the IT
and OT stuff because IT plans something.
There is something made in the ERP system.
There is a project being done.
And then in the factory,
something happens.
And how do you bring those two together?
That's I think where actually domain
knowledge comes to being right.
It's not just what happens on the floor.
It's just not what the
schedule you need both.
Denis Gontcharov: Yes.
I think by your expression of you
need both is very relevant here.
Let's focus a discussion on a
concrete use case of, for example,
building a process dashboard.
In my experience, the main problem
has been that the domain experts,
they know what they need, but they
are not able to build it because it
requires very strong digital skills.
Whereas the IT people, they
know how to build it, but they
don't have the knowledge to
understand what they have to build.
So you have these two different teams
and with different skills, and they both
have to be aligned to create something
that generates value for the business.
How would you approach this problem?
Luke van Enkhuizen: Yeah, I think
this is a, it's a perfect example
because you can have that dashboard
built very quickly, if you have already
all the information in place, if you
have all your ducks in a row, you
know exactly what needs to be done.
It's relatively easy to do this, but
if you don't, then you need to rely
on a lot of knowledge that you have
to be acquired, whether experience
from a consultant or from a software
that is a pre built solution, or
you have to make something yourself.
And I think that's where the distinction
is going to lie what you're going
to do and how long it will take.
Denis Gontcharov: Would you say that
domain experts always know what they need?
An additional controversial stance on top
of our earlier mention of having only an
AI in the factory is that you could argue
that in the times of Henry Ford, when he
developed the automobile, the people were
asking for faster horses, not automobiles.
So aren't we constraining ourselves
by only taking into account the
opinion of the main experts?
Luke van Enkhuizen: Yeah.
Well, and this is a very interesting
because if that person is a someone that
is in charge of selling horses, obviously
they're going to sell you more horses.
And if you're going to
design an automobile.
You need to have like, you know,
a bit of a different approach.
And I think this is very similar to
what's happening in the industry, right?
You have large established players
that have fleets of software
consultants and solution architects
that pay for outlets to promote their
product as the one and only solution.
And it's faster, it's better, exactly
all these things, but it's not different.
It's just.
Bigger and larger and more complex.
And if you then try, if you decide
to completely turn it around and you
say, okay, no, we're going to go for
the approach of the unified namespace.
We're going to put all our data in a
structured way, see it as the first
class citizen approaches from a digital
first perspective and build that
foundation where all the knowledge
is going to go into, yeah, you will
find completely different answers.
And I think of course.
Domain knowledge is super important and
needs to be captured in various ways.
But it comes from both sides.
It comes from what you already have
created, what you already have built.
And even it people already in the
company already built things and the
probably that the operators don't
know about and vice versa, there
is already data in various places.
So it's really important
to bring that together.
Denis Gontcharov: I fully agree.
At the end of the day, the way
I see it is that both camps.
We'll have to learn some of
the language of the other camp.
So what I mean by this is that the
domain experts, they should be aware
of the new possibilities that are
possible with digital infrastructure,
like the unified namespace, but
they'd also with AI and conversely,
the IT people, they cannot just stay
in their realm of bits and bytes.
They also will have to do an effort
to better understand the needs.
Of the, the main experts to better
be able to sell or to tailor their
it solution to their problems.
Luke van Enkhuizen: Yeah,
definitely fully agree there.
I am also not saying myself a couple of
years back, I would also still preach
that ERP would be the way forward for
companies to make their central source
of truth, for example, and put all
the knowledge of their company in
there and me as a domain expert would
amplify that message from ERP companies
and now only recently something new has
to be coming in that shows you there
is another way for example, event
driven architecture and this allows
you to capture what's happening in the
factory as the events as they reoccur,
just like you would be looking in a
factory and seeing things happening
event by event and machine starts
and stops and order is launched.
An order is created changed,
all these things are events.
And so it is kind of like, I'm
in the domain of that factory
and I'm an industry expert.
That's what I mean with domain expertise.
But when I mean domain
knowledge, I mean more.
The knowledge that's in the
factory itself, like that you're
building up the experience.
for me, so that's more like the site
and the enterprise in my definition,
perhaps we can improve that a bit,
but so from that perspective there,
I think that if you looking at the
domain knowledge in your factory,,
what's going on, you can observe it.
Right.
And that's also the observability
is what makes everything different.
If you do it right.
If you record everything in
history, because then at a certain
point, all your knowledge will be.
Already there.
You just don't see it.
That's where AI can come in to process
vast amount of data into information
that you can take action upon.
So I think that's the major shift
you can have a horse that runs
faster, but you can have a car
that's built completely different.
And there's a chemical process
going on completely different
than what's happening in a horse.
In the end of the day, they both
move you forward, but the car will
do it much faster, much further,
has much more growth potential.
And so forth.
Denis Gontcharov: Yeah.
I like your distinction between domain
expertise, something that the consultant
should have to empathize and to understand
the domain expert who himself or herself
has the very deep domain knowledge.
I think by you having this expertise,
that way you realize that, hold on.
I shouldn't sell horses.
I shouldn't be the ERP expert.
Instead, I should really think
about what's good for my industry.
And that, I guess, made you realize
that no, actually the unified
namespace would be a better candidate
to hold all the data centrally.
Luke van Enkhuizen: Yeah.
I've been converted to
that religion as well.
Sometimes I joke about it.
And this is so relevant for that
future, that story that you said in
the beginning about the factory of
the future where it runs automated.
I only think the only way how
this is possible is that events
are recorded as they happen.
They are showing you a pattern, even
though you don't see it yourself.
It's like the stock
market in the beginning.
I can, you know, we can both look
at charts and use software to make
some visualizations, but on a larger
scale, you need more than that.
You need multiple layers of calculations
and complexities to see what's the
market trend is really going right.
And so the same goes for your factory.
So it, it will need all that data
from years of history to predict
some of the future because the
history doesn't always repeat.
But it, it rhymes, right?
And, and that's basically where then if
you have a long term at a time, start
capturing all these events and use that
to then control the process with as well.
If a state changes or a calculation
changes, then make a process decision
to change something in the future.
That is where I think
the real magic happened.
But the major distinction
is that it's real time.
And it's based on historical data.
The historian is in place.
And instead of something that is only
responding to the data in a moment.
Like if you look at it, for
example, a swimming pool, the
water level in the swimming pool
is always leveled the same way.
How is that possible?
If somebody jumps in and water leaks out,
that's happens with a very simple process
that just checks the water level, right?
That's a feedback loop as well in itself.
But it doesn't record how much that
water level has changed over time.
And it doesn't allow you to see
what's happening around it, not
only with the pool, but the whole,
the whole building itself and how
many people are buying tickets for
the pool or something like that.
So it's like it, it extends is way
beyond the boundaries of the moment.
Right.
And that's what you need for AI.
I think to make decisions upon.
Denis Gontcharov: Yeah, for sure.
I want to refocus back on the role of AI.
We mentioned earlier in the podcast
that there'll be a man and a dog.
I don't fully agree with this.
I think the promise of AI will be an
automation of a lot of decisions that
were usually made by domain knowledge,
but I don't think an engineer will
ever be able to be fully replaced For
the simple reason that AI relies on
data and the data is not always there.
We are limited by the sensors that we
can install, especially we don't have
a complete view, but what I think is
that AI will perhaps, it's hard to put
a number on it, but it will definitely
outsource a lot of the lower level
decisions that an engineer makes now.
The main knowledge will always
in that regard be important.
In the near future to design the
dashboards that we need so that to
find use case that create the value.
But even in the future, even when we
have an AI steering most of the process,
I think the engineer will always be
there to watch over the AI, assume
responsibility and perhaps manage the
most difficult, ambiguous decisions.
Do you tend to agree with this statement?
Luke van Enkhuizen: I'm not entirely sure.
I don't have enough domain
expertise for your specific metals
background in the aluminum industry.
For example, I don't have
the domain expertise.
So this is such a thing example.
I wouldn't know because I don't know
exactly is the, the domain knowledge that
I need to have the domain expertise for
your specific aluminum industry example.
So I would not be able to.
Judge the book by its cover and saying.
Oh, I think, yeah, sure.
It's aluminum.
It must be a continuous process.
Oh, it sounds like temperatures.
Sure we can do this for the
argument, but I don't know.
And I think anybody that will say that
to a company like saying, Oh, you know,
we are experts in digital transformation
we work with various industries.
We know just the solution for you.
I mean, I don't know.
There must be like having some
psychic abilities or something, but
I don't think they can know this.
This is where I think your expertise shows
for years in a factory for my background.
I can say, yes, I think.
A lot of processes around starting
manufacturing, logistics, self driving
vehicles, basically everything that is
somewhat predictable and can be simulated.
I think that's a very
important distinction as well.
Can you simulate it with the software?
So for example, a sheet
metal part can be bent.
You can simulate that.
Logistics, you can simulate that.
So if you can simulate it, then.
I think in the end, you can fully
automate it without humans, but
everything that's not in that realm
that you cannot really simulate
because there's too many variables.
I don't think you can, like,
that's where you need a lot more of
computing power that we are not at
yet, but everything depends on the
time frame we're looking at here.
Are we looking at next
year, five years, 10 years?
And then I cannot look so far in
the future because two years ago,
did we expect that AI tools like
just large language models would
be so powerful today already.
I didn't expect that.
I'm still blown away every
day that I work with them.
So, and seeing how this speeds up
like developments, I don't know.
It's hard to predict
the exponential growth.
Denis Gontcharov: Yeah, it's a good point.
I think we really have to be
think about where we apply the
AI to which games we let it play.
Okay.
I think the things you mentioned,
things like supply chain planning, those
are relatively deterministic, like a
game of chess and can be calculated.
Whereas if you look at, for example, in
aluminium production, the electrolysis
cell, the process that we have been
running for the last 200 years, almost.
We still don't really know what
happens chemically inside the cell.
So that it's essentially
still a very big black box.
We know that which levers we
pull, that aluminum comes out.
But we don't have a complete scientific
explanation of what happens there.
And I think that's also part of
the beauty of this process, that we
will always need someone with a gut
feeling to decide what will happen.
Because the cell may go out of control for
a variety of reasons, and it's not really
clear what's going on before the fact.
Luke van Enkhuizen: That's fascinating.
I had no idea of this.
I think, a perfect example of your
domain expertise that you have.
Let's uncover this a little
bit for just a few minutes.
What's, what's happening in an aluminum
smelter just for an example here.
And why is it so hard to fully automate?
Denis Gontcharov: Sure, so if we
imagine an electrolysis process to
produce aluminium, what we essentially
have is a big bottle, made of steel,
that is filled with molten salt.
So it's essentially a very toxic
mixture of ions and in which we
slowly dissolve aluminium oxide,
just aluminium bound to the oxygen.
And the name of the game is to
rip the oxygen from the aluminium
so that you get pure aluminium.
Now, the way this dissolution happens
of the aluminium oxide in the, what we
call the cryolite molten salt is unknown.
We are not sure which molecules
or which ions are forming.
are, we have some idea.
The only thing we know is that we heat
it to a 960 degree temperature and pump a
large electric current through it, but we
get aluminum at the bottom and CO2 that
evaporates, but that's essentially it.
We have some understanding of the
physics and chemistry, but not more.
Then there's also a lot of other
effects such as magnetic waves because
of the very large electric current.
You have the heat losses, and then you
have all the bath chemistry that changes.
Essentially, you're steering the
process with just a few levers.
You're changing the voltage, you can
add some pepper and salt, you can
change the distance of the anodes and
the cathodes, you observe the bath
chemistry by measuring only once a day,
and this is where the problem occurs.
Because it's such a harsh environment,
it's so dusty, there's strong magnetic
fields, there's a high temperature,
very toxic chemistry, is that you can
only measure, and also don't forget
you have not one pot, but you have up
to hundreds of pots in one factory.
So there's like no way you can observe
all of them in complete detail if you
have to make , manual measurements.
So what you do is you
essentially do the best you can.
With a very few information that you have.
Luke van Enkhuizen: What kind
of data points do they normally
collect in a process like this?
Denis Gontcharov: So the information
you have on a continuous scale
is the voltage and the current,
the temperature of the cell.
You have like once per day, if a
team goes and measure it, you also
tend to measure the bot chemistry.
Also every, let's say 32 hours
up to 34 hours, you can also
measure the composition of the
metal to see how pure it is.
You can also measure certain
voltage drops to see how much.
of the electrical energy you
lose in your connections.
But that's really about it.
You don't really have
that much to work it.
So you have, let's say, a dozen
of input variables, which you
have to steer a physical process
that you barely understand.
It's a bit like cooking with experience.
You're cooking the steak, you're
adding something, but you're not
really sure of the exact chemistry.
But you don't have to as a cook.
If you're an experienced cook
like Gordon Ramsay, you can cook
the perfect steak without knowing
anything about the chemistry.
Luke van Enkhuizen: Right.
Because the compound of the steak,
the meat itself could be a little
bit different, the butter could
be a little bit different, the
environmental conditions, you
observe the changes, what you're
seeing visually, and you are perhaps
smelling it whatever, all these things.
But so why are you, are you then limited
to collecting data points only once a
day by checking into these spots what
limits, for example The, operators
from capturing these data points every
second or every millisecond, even.
Denis Gontcharov: The number of pots
you have in a hundred of electrolysis
spots, bottoms in a single factory.
You only have a team of, let's say
five people that are responsible for
the measurements and they can only
do so many measurements per day.
Physically.
You can't use robots or fully
automate this process because
of the strong magnetic fields
that damages the equipment.
You also have the very high
temperature and the dirty environments,
but of course, also the cost.
Imagine even if you could develop a fully
autonomous robot, you would which some
plants are actually developing right now.
The benefit still has
to outweigh the costs.
Luke van Enkhuizen: And
so let me get this right.
So you cannot place normal sensors,
even industrial grade sensors in this
process on various data point that are
capturing continuously because of the
magnetic fields, or do I get this wrong?
So a normal temperature sensor.
Denis Gontcharov: Partly the
magnetic fields, partly the
heat, but also the costs.
There are always been projects where they
want to retrofit existing cells by, for
example, adding a mesh around each anode
block through which the current flows.
This would give you exact
information about how much
electricity goes into the bath.
From every anode and that will
give you very useful information.
Problem is that retrofitting all of
your hundreds of bots with this sensor
technology is extremely expensive.
And the process itself, just by
having 200 years of experience is
already about 95 percent optimal.
So the question is with a few additional
percentage points of more efficiency
that you gain with this technology,
would you cover the cost of investment?
And that question is still
being debated to this day
Luke van Enkhuizen: That's a very
interesting topic to emphasize.
It's sort of decision on when, when
it makes sense to invest into further
collection of data points depends highly
on how accurate your current process is.
And again, the domain knowledge
from 200 years of experience,
if this makes sense or not.
And so there will be people on the
shop floor saying, we got this,
we're doing this for a whole career.
We know how it's like, we just
want it a little bit better.
A little bit faster, maybe, and
then there might be people like us
coming from the outside or anybody
else saying, where's the data?
Give me more data at all costs because
data is the most important thing because
that will show us things we don't even
know that are there in the first place.
How I observe it a little bit here,
like not saying one of the two is right.
I'm just like, it's a difference
between approaches, I guess.
Denis Gontcharov: for sure.
But data is important.
For example, an area where
I see a lot of potential for
improvement is not necessarily a
better steering of the process.
That's arguably already good enough.
It's more when people do operations
on the pot, when they replace an anode
or when they start up a new cell, the
way people like manual, whenever a
manual decision is made by an operator,
that's where you introduce variability.
So a big trend in the industry right
now is trying to record how well
a certain operation on the pot has
been performed to then potentially
identify problems before they occur.
Luke van Enkhuizen: So there's this
predictive maintenance in this, this
context here, but in a certain way, right.
Denis Gontcharov: Yeah,
it can be compared to it.
It's more about checking like verification
of the operations made by a human,
Luke van Enkhuizen: Right.
But for this, you also need to capture
a lot of data then to, to make these
kinds of predictions, I assume.
So in most smelters, do they already
capture historical data in a somewhat
structured manner about starting
and stop and those kinds of things?
Denis Gontcharov: mostly on paper.
One example I've seen which really
impressed me was that a lot of
operations are done by a crane.
That moves across the pots.
So above them in the air, what they have
did is that they mounted a camera on the
operator seat that records the operation.
And, but that's where you also have
in your MES system, the start and
the stop of a certain operation,
because after the press of a button,
I am now changing anode one, two,
three, the camera starts recording.
And then when he clicks on,
I finished my operation.
You have all the information you need.
You have the start and the
end of his operation and you
have all the video material.
That is being analyzed by an AI to gauge
how well the operation has been done.
But that's something
relatively new in the industry.
Luke van Enkhuizen: Yeah.
Vision systems.
I was actually already thinking about this
in the beginning when you told me about
observing if a process is going well.
I was like, what can you do with vision?
What can you do with sensors?
What can you do even with sounds
or even measuring magnetic fields?
There are probably ways that you could in
a certain way, measure all these things.
And they should give you some hints and
insights how well things are going and
where improvement lies and the effective
OE and which again, you will need if you
didn't start bringing in outside factors.
I think in a previous
episode, you mentioned energy
consumption and optimization.
And I think this is huge that you are
strategically can time your maintenance on
times where the energy grid requires you
to do so or predict even when you should
do this based on the energy grid usage,
just thinking out of the box here a bit,
but I think those are kind of the things
that you could do if you set it up, right.
So I see it from an outsider, but
you are the domain expert here.
I'm just asking you now
how you feel about this.
Denis Gontcharov: Well, I'm not
a domain expert, so to speak.
I have some domain knowledge of
the process, but I think that's
actually a great point to tie it
back to the discussion about the
importance of domain knowledge.
How can you tell if an anode has
been changed correctly or not?
With my knowledge, I can just
realize that this is important.
It's important to ask, but then an
actual domain expert, which would
be like the engineer or the operator
working there for over a decade.
We need the information.
We need the domain knowledge of those
people to be able to Label the subsequent
AI data, the visual data that we generate.
Luke van Enkhuizen: Yes, indeed.
So that would be a perfect example in it.
So how does a good change look like
and how is that reflected in the data
in other places down in the process?
So for example, if you make that
change, did anything else change
in measures right after that?
Something that you might have not noticed.
You know, like there could be so
many factors that are related that
we just don't see with the naked eye.
I think it's good example because when
we go back to the beginning, we'll be
just a factory with a man in a dog.
There is potential for it, I
feel, but I think there's a
long way to go with robotics and
just the physical limitations.
But it surprises me a little bit that
a human can do it and not a robot, but
it has to do with some environments.
But then are these really like the first
principles that you can really not change?
I somehow feel that if we find
a way to overcome these things,
then it's just a cost question.
And if it's a cost question, then I
think we should include a lot more
than just what's happening inside
the four walls, but also externally.
As for example, I mentioned energy
or regulations or CO2 all kinds of
things you can bring in or supply chain
optimization to make the decision a
bit more a bit more founded in data.
Denis Gontcharov: Let's focus
on your question of, do you
need a human for this decision?
And I think it's a very
important question.
And the answer is yes or no.
If you train the vision model to recognize
if NLs are placed correctly or not.
I'm a strong believer that this can
be completely done by a machine.
If you label enough images, there's
only so many different variables.
If you look at the way.
You see the slack.
If there's lots of light,
it means you have holes.
That's the thing a machine can answer.
But if you go to the more higher level
questions, for example, you can actually
compare an electrolysis to a human.
if it gets too warm, temperature
above 1000 degrees we actually
say in the, in the industry that
both has a fever, it's sick.
The question is, why is it sick?
Why is it, why is the temperature high?
Why does it have a fever?
And that question is extremely difficult
to answer, even for an experienced
engineer with over a decade of experience.
And that's something that I
do not see being automated
anytime soon, by an algorithm.
Luke van Enkhuizen: right.
Then we are at a point where we
just are managing by exception.
And I think that's
already a massive progress
And if you get sick there's
probably a very complicated process.
Like if humans get sick, we
don't always really know what
the accurate diagnosis is.
We have a lot of ideas.
We can look for viruses, but we
have to probably do something
to really check in the body.
Look at the blood, look at the saliva
to know what's really going on.
You can not just like talk to a computer
and a computer tell you with accuracy.
What you have, right?
So I think that that is a great example
where you are looking at the exceptions.
But I feel even with that part
getting sick, with the diagnosis
of that part getting sick, you
could probably also use data too.
Have a extra pair of eyes with
you, because what would this,
the specialist do to research
this, which steps is he taking?
What does he see?
What does he hear?
What does he note?
What does he see in the historical data?
if you look all the, bring all those
things together, that that's actually
the domain knowledge that he has,
and that can be recorded some way.
And it would be used as a blueprint
for another time, something gets sick.
And if you record this often
enough, you start to see a pattern
and then you could probably
predict it even before it happens.
Denis Gontcharov: Yeah, I
think no one can really tell.
It's a very exciting period in
history where we'll see just how
smart the systems can become.
My gut feeling I wonder what your opinion
on this is that I think we may see our
dear Pareto appear as well in these
discussions where, for instance, 80
percent of the decisions currently made by
an experienced engineer could be done just
as well by an AI, and we will have the 20
percent of the most difficult decisions
will remain in the hands of the engineer.
Luke van Enkhuizen: Yeah.
I think this is something we will
see across all industries that a lot
of workload will be offloaded to AI
systems where the humans are monitoring.
It's kind of like an
air traffic controller.
The planes can perfectly take
off and land by themselves.
You know you can perfectly schedule
which plane lands where automated, but
there is still traffic control necessary
and quite a few of them because what
if this computer is wrong and what if
something unplanned happens or what
if and last minute change is required.
And so I don't think we will get to
the point that there will be a man and
a dog, that the dog will start barking
if the man wants to touch the button.
I think we're far from there indeed
in this current time and place.
But I do think we need to move forward
very quickly to a world where a lot of
this decision making can be assisted or
at least suggested by AI systems very
quickly where it will predict like,
hey, you know, check this out,
probably something wrong here.
Before you actually already noticed it,
like like when it was already too late,
I think this is a great example for that.
Yeah, I think this is a great
thing to wrap this episode
almost up on this topic already.
Cause we talked about the
importance of domain knowledge.
Of course, there's much more to be
said about how this will affect for
a company to want to do a project,
but just in covering that really,
that, that's really foundational,
you know view on domain expertise.
I think it's quite important.
Denis Gontcharov: Yes, I fully agree.
And in a sense, if you want to
transport your domain knowledge into
an AI, the precondition you need to
fulfill is that you actually need the
clean data, which again, highlights
the importance of investing into
a proper digital infrastructure to
capture all the data from your systems.
Luke van Enkhuizen: Yeah.
And then you'll quite quickly notice
that to then do something with that
data becomes more accessible to anybody
in a factory, if you do this right.
And so I think we will uncover
more about the unified namespace
and the approach behind this.
For example, you want to make a
dashboard in the next episodes because
that will be more specific topic
of like how to build a quick
proof of concept or a solution.
And what do you need by that?
I think that's better for what
next episode to talk about.
What do you think?
Denis Gontcharov: Yeah, I fully agree.
And I really look forward
to that discussion.
Luke van Enkhuizen: Well,
that was this week episode.
Thanks for listening.
Bye bye.
Denis Gontcharov: Bye bye.