Introduction

So great, thanks all for being here.

It's really cool to stand here in this very nice church.

It's not every day that you get to speak.

Some people speak in church a lot, but I don't.

So my name is Rian Onof.

I'm currently IDSO 2030 Transformation Manager in Shell.

That's a lot and I'm not going to explain exactly what my role is because I'm not here necessarily on behalf of Shell.

I just wanted to share with you some of my experiences and insights and some of my ideas of different uses of AI that I've seen and how we can relate it.

And not so much Gen AI, more generic, more other uses as well.

Predictive Maintenance on Physical Assets

A North Sea Platform Scenario: When a Critical Pump Fails

So I am going to start with a very Shelley picture.

So imagine you are not in a beautiful church like this, but in an oil and gas platform on the North Sea.

This facility is giving us energy, bringing energy to our households and gives us independence from

other countries like Russia or the US or other areas where there might be war at the moment

and it's been running for decades.

It's full of complex machinery that somehow still works

and we have shift crews up here in a two weeks on two weeks off rotation

that are keeping it all safe and efficient

and keep the uptime high, right?

And then suddenly what happens?

There is this pump and it's actually broken.

It doesn't work anymore.

So what do we do now?

Yeah, the whole facility can't work

because this is a critical pump.

We really need this one to operate our facilities.

And yeah, you can't just replace it.

I mean, maybe we need a helicopter to fly a new one in if it's even on stock so quickly.

So it could take days to repair millions of dollars per day and it interrupts our energy supply.

We don't like that.

So what do you normally do?

We have time -based maintenance, which means this type of equipment, we repair it every two years, for instance.

Well, that's risky because it could break earlier than two years.

and it's also costly because maybe it could run for 10 years actually but we've already replaced

it so what do we do yeah there's no alternative because you can't predict the future you don't

know when it will break right but with AI we can we can predict the future and know exactly when

our equipment will fill so how do we do it we use sensors because before a pump fails it already

starts behaving erratically the the vibrations are different it vibrates

higher or lower or infrequently and yeah so we know that with sensors you it's

difficult to tell but if you use the sensor data from past failures and you

give that to machine learning AI can interpret that and know for the future

when will the equipment fill so this is the cost deck right maybe I can do with

with my mouse.

I can do it in person.

Okay.

From Sensors to AI: The Technology Stack

This is the technology stack.

So here, this is the

sensor itself, right?

It sensors the vibrations, and it monitors actually what's happening in the

telemetry that's coming in.

Then it sends that to our LoRaWAN wireless connector.

So this is what's

connected to the Wi -Fi of our site.

And from there, it goes to our secure cloud, where we interpret

the data we clean it up and we store it and that's where the AI compares it with

historical data and knows the live feed of hey what is really going on now if

you are an expert you can immediately see which of these two is the healthy

one yeah icons but our AI is impeccable in that situation so you might even get

a whatsapp message or at Emma's team school or your maintenance system I take

this into account and make sure you fix your stuff so that's integrated in our

in our processes.

Scaling Predictive Maintenance Across Equipment

Now, I mentioned the pump, I mentioned one pump in one platform, but there's

lots of different equipments you can do this with, right?

Motors and compressors and everything.

So, for instance, in a site in Pernis, we have 4 ,000 pieces of equipment that have this sensor

and have this type of predictive maintenance.

And in the past, this idea is not new.

We had

this 10 years ago.

But what's new is that the price of these sensors, the price of the LoRaWAN

connectivity and the reliability and price of the AI interpreting it, is so much better.

It's

stronger performance, lower price, that we can roll it out at scale.

Applying the Same Idea to IT Operations

Now, I promise you to link this

also to our IT landscapes, because the clue is we can use the same idea of predictive maintenance

from our physical assets to our IT assets.

We run the service, the networks,

the phones, everything that we do and our applications.

We can apply the same

concept and predict before things fail.

IT Support Maturity: From Reactive Monitoring to Prediction

So what does it look like?

If you think

of an IT support desk, right?

In the very traditional way of working, this

is a level one maturity.

These guys do not have predictive

maintenance.

What do we have?

We have an angry user on the phone because the

system is down.

That's the first time the IT department actually hears about it

and this is what happens, the service degrades and it is actually out, it's not

working.

Only when they hear about it they can start fixing it, right?

Not ideal.

Can you hear me?

Okay.

I think it's not a very big room but okay.

So if you go to

the level two with the enhanced monitoring you don't need to wait for the user to call you can

already notify it with your mode monitoring you see when the system is down and then the

i .t department can start fixing it earlier but the real opportunity is here even before it goes down

the system already starts behaving erratically just like the pump that is vibrating yeah so

So whether it's network switches that are overloading and stacks that are high

or things that are erratic behavior that we see, this is what you can notice.

And it's the same AI principles of noticing, observing your system's behavior

and having AI interpret that and predict what's going to happen.

The IT “Superpower”: Automated Fixes with AI Agents

And here comes the superpower of IT.

Because in our physical assets, if you want to replace the pump,

you still need to fly that helicopter out, right?

in IT you don't have to have any person fixing it because quite often a fix is resetting a switch

or stopping a task or running a script that you already knew and it's the AI agents that can run

that you can have AI agents monitoring your landscape and fixing it so in that case what

you can have is that instead of being called in 3 am by a help desk employee escalating a huge issue

you actually wake up in the morning with a notification a high priority incident was

predicted and fixed before the users even noticed it and that's what we see in this this level three

is that before things start to behave weirdly we already fix it and that's

what we're working on we see that starting in some areas but it's really

the future that you don't have to work fix these things in person you can

predict it before it happens now we also have level four and five and this is

Beyond Incidents: Predicting Demand and Scaling Preemptively

where you you go beyond the observing the erratic behavior but you already predict before the

erratic behavior happens so for instance in my company we all get our bonus payments are listed

on the website on 16th of february so i know that 16th of february early in the morning there's a

huge spike in usage of that website because everyone wants to know what their bonus is this

year right now as a human i can predict that and we can scale up for that but there's other

behaviors in our systems that we are it's difficult for us to be hit to

predict but AI can predict it so there's multiple updates that are running at the

same time multiple batch jobs that are happening maybe people log on in the

morning at the same time maybe something happens in a another country where there

is a holiday there for some kind of behavior is high on the network so in

these cases our AI can predict the usage of our if our applications and systems

and already scale up preemptively for these things and that way our landscape is not only

reacting to incidents predicting incidents and fixing them but it's also optimizing the entire

landscape based on what's already happening and this is already possible today it's just a matter

of linking all these different elements together to implement it and that's where it gets hard

actually because if you have if you have your network components you need to

match it to what is it doing in the business to really understand what

Linking IT Components to Business Impact

you're doing and to imply that intelligence to it so here's an example

of a network switch that is an outage our AI is predicting that it's gonna

fail but what is this network switch doing is it supporting an office at so

So imagine it's 8 p .m.

on a Friday.

Our office in The Hague is completely empty.

So an empty office floor with no internet is not a very big problem.

It could also be the traders' floor in New York,

which are trading very heavily at this moment

and actually, unfortunately, making a lot of profit

from the instability in the world.

But yeah, having an outage there for a couple of minutes is not great.

It could also be the German Autobahn,

where everyone who's just going on ski holiday

day is trying to get gas and having our point of sale offline at that moment is not really ideal

either so it's about mapping our entire landscape to what is does it mean to the business and what

are we doing with it because that way we can predict what we're doing with it and to do that

we need the operators that are currently operating that landscape and we need to have them help us

make that mapping in order to apply this logic to it but then we also get to the really difficult

The People Side: Roles, Upskilling, and Adoption Challenges

part and the really difficult part is exactly those people that we need the help of to implement

all these brilliant ideas are the same people who are looking at this AI with a wary eyes because

the operators that are currently supporting the system they benefit from a better system that

doesn't wake them up in the middle of the night right no calls at 3am no constant firefighting

more professional work happier users who doesn't want that well these users that might see their

own jobs evaporate today what is my role then in the future as an operator am i actually capable

of doing that role how much upscaling is needed will there still be jobs i have a team of 10

people now will it be five in the future or zero for this so what we need to do is paint the

picture of this future landscape understand what roles do we have then in that future and have give

people a feeling of hey i fit in there and i want to do this and i want to learn the skills to do

that and and help them upskill to that future um but yeah i think you can talk for a whole day of

conference about this people angle as well and i wanted to um to leave it there and just give you

Conclusion

this overview of the idea of predictive analytics for assets we can apply that as well for the

predictive analytics for rnt landscapes going to not only predicting problems also solving them

having potentially autonomous systems and agents that apply the fixes even before something breaks

and then go from there to make sure we have a smoothly running landscape overall i think that

is something that a lot of companies can use and whether you're big or small if you can make your

systems run more smoothly you will always have happier customers thank you we only have one mic

Q&A

so uh let's do as good as we can uh thank you first of all well the boss was so thank you

question or remarks of people with regard to this presentation how much you are allowed to answer

Tools and Platforms Used at Shell (High Level)

with this but as far as you can i'll be interested what tools at shell make this happen so what are

the main ai products tools models that you used to to build this that you show us yeah so to be

honest for some of these i only know the shell internal term of the project that we have so i

think we do a lot of the market standards the bigger market standard tools that we have and

And even ServiceNow we use for our operations has a lot of AI also embedded in it.

And the Dynatrace observability is what we use as one of the core components as well.

Some of the agents, we're still using multiple tools and trying to figure out which ones are the best and working with different pilots.

I could look up the details, but then I wouldn't be able to share.

So I'll leave it at this high level and then...

Cost, Monitoring Strategy, and Prioritizing What Matters

Yeah, so there's a number of elements to the cost of the monitoring, right?

So first of all, for some reason, we identify the difference between full stack monitoring

and all and basic monitoring.

And my colleague explained to me in length the difference.

But the fact is, the basic monitoring was a percentage of the cost from the full stack,

and was achieving 80 or 90 % of the results.

So for every solution, and this changes over maturity, you have to look at how much monitoring is needed to achieve your business purpose.

But I think one of the main costs in the past of the monitoring was, well, if you monitor something, you have to do something with the results, right?

Like, is the person then looking at this monitoring?

Is they then taking action?

And this is where the AI really helps, because you don't have to have a person looking at the results.

you can have agents interpreting all this data so there's loads and loads of data so

it takes a lot of time to interpret it but ai is really good at that and taking the key points of

it and taking the actions and agents can help you take that action but we are we are looking at ways

to i think it's becoming more and more affordable and much cheaper to do a lot of the monitoring

and there are more and more options to do that however we did start and that's why the mapping

to the business processes are as important so we started with the critical business processes

so for instance trading it's end of day for hrs the payroll payments for for downstream is the

point of sales in in the key like the keep if you have a gas station at the autobahn it's important

if you have a tiny one in north groningen somewhere it's not so important right so the bigger ones

that have more sales they're more important so if you map it to the business process you can see

which are most important and then you can focus your monitoring on those you can also choose to

focus on the ones that used to break a lot so sometimes our business critical business processes

are are already the most reliable while there's others that are less reliable and they tend to

break a lot and then we choose to focus the monitoring more on those but yeah it's always

a cost -benefit trade -off that changes a lot over time as the technology increases and gets cheaper

the trade -off changes yeah I know that's the ambition so so especially the the

agent equal to actually implement it yeah I know that's more division but the

first steps of the monitoring up to what we call level three we do have

have implemented for key systems.

So we have our whole landscape are,

some are just level one because they're not that important.

And some are level two, level three.

And for all those division is to get to level five,

but they're not there yet.

Yeah.

So it's a mix.

Over there.

Validating the Mapping Between Systems and Business Processes

Hi, good afternoon.

How do you validate the mapping

of the departments correctly done by the operators?

Yeah, so we had a team set out a process

where we have like now excel comes back i'm sorry but excel just an excel file with like a list of

which are the elements we think maps to this process and it's all based on our service now

mapping with the with the ids but then which how do they map to a business process yeah we had to

put a process around that and how do you validate that well we trust them a little bit but then we

check it of course in the reports and we check the data and we make sure it matches but it's

and then when there's something incorrect you you fix it but we had like for each business

process then a full dashboard so where you see all the elements that contribute to it

and you see then once you have it you can easily see that it's correct because there's a spike

somewhere and it's the user that complaining that matches with what there is so it is an ongoing

optimization and sometimes if there's a mistake you have to sort of fix it along the way

but there is a there is a process of asking information checking the information running

the reports behind it to make sure it fits and then optimizing as you go yeah but it's

Could This Work for Healthcare? Prediction, Data, and Privacy

it is a challenge yeah yeah a question that might be a step over to the moment perspective

i thought about would it be possible to apply this kind of monitoring system and interpretation

system to, for example, the medical data that are installed on the computers of our medics,

which will tell us what the blood consistency was of a person who had three tests in 10

years, and if you combine that to a national system, this data, that might come up with

with a prediction element as to which diseases

will kick in first at which point.

Then the second question is,

how would you safeguard the data there?

I mean, in your case, you need to safeguard that

because of other firms working with the same kind of system.

But in the other case, we'd have to safeguard data

of personal interest.

So, in theory, what, question to you, I think, technically it would be possible, but what

would you think about going this way, thinking about this, let's say having a monitoring

system on health, not globally, let's say nationally, combining all the data from our

are medics and their patients in order to predict

who might be the first to have either a heart attack

or another kind of a disease.

What would you think about this?

I want to give a response to that, though.

Sure.

Yeah, so first of all, I think we use, of course,

AI interpretation of data in medical field a lot,

but usually on individual cases.

So scans are interpreted by AI and by human doctors,

and then you see that the AI is more accurate.

If you want to do this meta view,

of course, the whole challenge of data privacy comes in there.

Who's allowed to do that?

There's also the interesting part of data training.

So in order to get this overarching knowledge

that the AI has, this machine learning we use,

we have to give them lots of data.

And we actually see that for some of these physical assets,

we were unable to apply it

because we did not have enough historical sensor data

for that type of machinery so it didn't know when it would break so some machinery in the past was

never measured so then you don't have historical data of when it will break and if you always

repair it before it breaks then you don't have that data so you should also have a lot of data

of humans then that get ill in order to protect predict when they get ill so i think that is

i think we don't have enough measurement because we usually only measure people if they get ill

and not if they're healthy so for things like the bevolkingsonderzoek maybe interesting yeah

ethically very difficult i think okay yeah space is moving there because more and more

there's talk about wearables that and not that i'm an expert on the subject but there are talks

about wearables that we monitor you constantly but of course it's your own decision to wear it

or not and then you have to figure out where does this data go which is also not so clear

but if the space is definitely moving to that and then it's the question of where you could go do

you draw the line and also if you need data then you need to let people get sick to see how they

go and then how do you choose which one if you get sick you you can say yeah there's a lot of

We have a lot of questions around us.

Anyone wants to add something?

Yeah, it's an interesting point, though, because there's almost more of an opportunity

for governmental organizations who champion health, like the Dutch, the health organizations

that champion overall health, like that do the bevolkingsonderzoek or that inform the

public about health risks and so on.

They might have more benefit of this than a pharmaceutical company which makes a profit

of selling drugs and it doesn't make a profit of healthy people so you would almost want maybe

yeah who who takes that accountability and i think it's maybe even the health insurance companies

that will benefit most from a healthy organization so maybe they're the ones that should invest in

this type of ai rather than the pharmaceutical companies probably depend on the country too

Yeah, I'm taking the Dutch context here, but sometimes I actually, no, now I'm digressed

completely.

I often think that that's what's wrong with the American health industry is that

their health insurance doesn't have as much power because they're the ones that benefit from a

healthy society and a low cost.

But yeah, so it's an interesting question for them.

Okay, thank you for your very interesting.

Do Sensors Really Pay Off? ROI and Process Change

I have some questions regarding the first talking about the sensor.

Since I'm in the industry of Pokemon to go to Europe, I'm curious about it.

so uh obviously putting the sensor is much cheaper than the helicopter yeah but i i'm not against

the body but i'd like to know how much it was beneficial to bring the sensor become i think

shell already has a lot of data about how much the pump runs and how much the motor runs when they

They build the whole wood rig for machinery.

I think all the devices are already good enough.

And then they design everything.

And then they have a regular maintenance plan

and have to stop to replace it.

So I think no sensor can prevent the extent of stop.

But I think if the item is already defective,

the oil rig should stop, no matter the sensor detected or the machine really stopped.

So I wonder how much it was really beneficial compared to changing using the regular internet

or actually happen of the stopping kinds of pump or any motor because I think there are multiple streams to keep the whole oil rig flow.

So for example, if there is one motor stop, one pump stop, I don't think the oil rig stops.

maybe there are alternative flow to running the whole, the factory.

So I wonder how much it was really beneficial to use in the sense.

Can you rephrase the question in two sentences?

Yeah, so no, it's actually a really good question because I really oversimplified the whole pump problem, right?

So we ask very good questions.

How much does the sensors really help?

Because can't you just use an alternative flow?

And aren't there other solutions to preventive maintenance?

Surely Shell already has good maintenance plans in place

and don't depend on the sensors, is what I'm hearing.

And the reality is that the uptime of our facilities

is not as high as it should be.

So you want 99 .9 % uptime.

when especially in our upstream facilities we don't hit that by at all we are too let's just

not say anything confidential but our our uptime is really too low so normally you have your plant

maintenance your your maintenance windows where you shut down your your site anyways because you

have to fix a lot of things and then you have your unplanned maintenance time that you have

to shut down to fix something which you didn't want to what you want is you want to have perfect

idea of your equipment health so that when you have your shutdown and for one week you're shutting

down and then you fix everything that needs to be fixed and then you start running again

so before that they have already a year of planning of what are we going to do in that one

week you want to have perfect understanding of which of my pumps need to be replaced

so that is one of the ways that it really helps us to optimize that maintenance window we do at

the plant shutdown and second when it comes to pumps you usually have two that are working

and one that is running and the other is the backup so that if it breaks you use the backup

and then you you fix the first one it does happen that they both break and then you really do have

a problem and you have to shut down so we have been using these sensors for our critical equipment

already more than 10 years and it has already shown very very significant

value for money.

The return of investment of these solutions is really high.

We are

now working in some of our new plans that we're building.

We are moving away

completely from time -based maintenance and only predictive maintenance to the

level that operators don't walk their rounds to do the usual checks every

week they just look at the sensors so some of the newer equipments really rely

on the sensors and can work much more efficiently also which much smaller

crews so this depends a bit on on which ones but we do see very high benefit for

these things one of the problems is you do have to be very clear on what is your

maintenance process and how are you changing your maintenance process with

this data so you can put the sensor on and change nothing else you don't get any profit you also

have to change your maintenance philosophy you have to think about which equipment you have to

change your planning for your maintenance turnarounds so all these elements have to

be in line with this new vision and that's why it's hard but yeah we see a lot of profit from it