2025 - AVEVA World - San Francisco - Process Industries (Chemicals, MMM, Pulp/Paper)
Methanex: Unlocking AI’s Potential with PI System Data & CONNECT: A Simplified, Scalable Approach
We use AVEVA PI System as our standard historian, capturing vast amounts of operational data across six countries to drive plant efficiency. Our goal is to integrate this data with foundational AI models hosted on Azure AI and models developed using Python; and generate executive reports via Power BI. The main challenge lies in securely extracting AVEVA PI System data from DMZs, made more complex by our global footprint. Existing solutions like PI Integrator and PI Web API require significant infrastructure, deep technical knowledge and lack a central data-sharing hub, making them non-scalable solutions. CONNECT addresses these issues by providing a lightweight, web-services-based platform for secure data transfer. This streamlined approach reduces integration costs and complexity, enabling rapid deployment of AI models and delivering real-time insights to leadership. As a result, we have accelerated data-driven decision-making and realized tangible benefits in plant performance.
Industry
Chemicals
Company
Methanex
Speaker
Julio Figueroa
Julio Figueroa is an innovative IT professional and IT Business Relationship Manager at Methanex, driving digital transformation and AI adoption. With a Master’s in Information Management, he specializes in leveraging AI for knowledge management, optimizing data-driven decision-making. Passionate about the intersection of AI and business strategy, he focuses on scalable solutions that enhance efficiency and innovation. Julio is eager to collaborate with like-minded professionals to drive impactful AI-driven change.
Company
Methanex
Speaker
Enrico Branca
Franco Branca is a Senior Process and Project engineer at Methanex, with over 16 years of expertise in hydrocarbon and chemical processing, capital project management, process optimization, and analytics. Trained as a Chemical Engineer, he excels in transforming complex challenges into innovative, efficient solutions that drive industry advancement. Passionate about bridging the gap between manufacturing and IT, Franco leverages technology to integrate robust systems with operational processes, enhancing performance and sustainability.
Session Code
SESS-239
Transcript
Thank you for being here.
My name is Franco, and Julio with me to you today. And today we're going to take you through our journey and some of the work we're doing in unlocking AI's potential with the PI data we have. And it's what we call a simplified and scalable approach. And hopefully, you'll find it interesting. You know, there'll be some things you'll be able to take away from it. As you know, these journeys are typically multidisciplinary, like Steven has mentioned. And we had to bring together operational and engineering experience with our data science and IT backgrounds to really solve the problems we face across our methanol production network.
So if you're unfamiliar with Methanex, don't know who we are, We are the world's largest producer and supplier of methanol. And we produce about ten point six million tons of methanol annually. We're based in eleven countries with six production sites.
And we have about three point seven billion dollars in revenue yearly. And we also have an integrated global supply chain with our own wholly owned waterfront shipping subsidiary.
But we're a small company. We only have about a thousand four hundred employees, yet we've got a global impact.
So if you're wondering what methanol is, and there's quite a few people that have asked me, at this event, it's a versatile, chemical that's biodegradable and it's used in everything from plastics to paints to fuels and adhesives. So it's very likely that you're holding a device right now or there's one in your pocket and that that has been built with methanol derived materials in it.
So I haven't got a slide of it up here, but our typical process relies on methane as a feedstock which is reformed at high temperatures to break it up into the building blocks that we use to make methanol molecules, hydrogen and carbon monoxide. That is then built at a high pressure in a reactor, after which we purify it in the distillation step, we store it and we ship it globally.
So ultimately, that produces a clean burning, cost competitive alternative fuel that, as you've heard, is an essential ingredient for everyday life.
So as I mentioned, we're a global company, And each one of our six production sites already has AVEVA PI Historian, AVEVA Asset Framework, and AVEVA Connect enabled. So that's a good starting point to give you some background.
And like many manufacturers, we face some common problems.
Of the three main ones I want to highlight, firstly is we have global experts sitting in different locations. So we need to be able to supply the right information to these guys to help the different sites. So we need that global collaboration.
Secondly, equipment failures can result in significant losses for our production. So we also want to utilize the data that we have that's untapped to predict and prevent failures before they occur. And then lastly, operating efficiency, a common theme as well.
We want to optimize our operations and help the operators themselves using AI to assist them.
So yeah, that's today we're going to share a bit about that journey. I'm going to hand over to Julio who'll walk you some of some of the examples and ideas that we're working with today. Thanks, Julio.
Hello, everybody.
My name is Julio. I am the IT manager for the manufacturing site we have in Canada, which is much colder than here. So don't complain about the weather in front of me, please. This is rude. So as Franco mentioned, we have been using PI System for many years, decades, actually.
But it was just during the last year when we were able to standardize our entire Py platform.
We didn't do just for the sake of the technology, we have enabled the business to start focusing on standardizing business processes. And this is a source of pride, and we're not going to hide.
But, you know, we reach a moment where we need to get prepared for the next phase. And the next phase is unlocking insights from our data using AI.
But AI might mean different things for different people. So let's talk about the AI promise.
This this place where everything will be easier, faster, and beautiful. You know? But that's true. And what is also true that we also have clean operational data in our PIE system. But how to connect these two words?
Connect is a hint. So we're not talking during this presentation about cultural aspects like getting support from the upper management, building multidisciplinary teams, or building the internal, skill set that you require. Let's talk about infrastructure. And at Methanex, we have found that Connect has been the solution that has allowed us to bridge these two words in a secure way, but also in a flexible way.
So at this at this point of the conference, I believe everyone knows about Connect. So we all know that it's a central hub where we can share data easily. You know? At Methodex, we are, using specific specifically the data service services model.
But let's talk about four use cases. During this presentation, we'll we'll share with you how we're using Connect to visualize data, how we are using foundational models with our PI data.
We'll show you a prototype of using PI data with AI agent. And finally, Franco will show you a real example of using ad hoc machine learning models with our data.
So let's start with visualizing data. This is even though it's not AI, this is critical because even though you can have great models or great analysis, you have to build visualizations that are appealing to your users. And at Methanex, we have mainly two kind of data consumers. One that are operational that I would say that are really well served with, PyVision, PyDataLink, but there are other kind of users that are they are strategic, you know, where they need data, consolidated data from different sources and in a different way, and this is where we are using Power BI.
What I would like to highlight from this experience is how much Connect has simplified our infrastructure.
Before having this standard, architecture, we one of the models that we had in order to achieve these reports in Power BI, it was a really heavy infrastructure.
PI Integrator or sometimes PI Web API, then SQL Server, our Microsoft Gateway, that's heavy. We have six manufacturing sites. When you scale that to six manufacturing sites, you can see how much effort is required to maintain this.
Once we migrated to Connect, this is what we have. Just a simple PI interface to connect, and this is how we build the Power BI dashboards.
This is how it looks when you scale it.
So Konnect has enabled us to deliver this kind of dashboard with a much lightweight IT infrastructure.
The second use case that we would like to share with you is about using foundational models. Okay? Just in case it's not a very familiar term for you, bear with bear with me two slides, two technical slides, very quick. So traditional machine learning approach is the following. Fall following with this example. Let's imagine you have a pump, and you want to predict a failure on the on your pump. What do you need to do?
You need to get a lot of clean data, then you train your model and do iterate until your model provides an accurate response.
If you have been involved in a project like this, you will know that this can be time consuming. Sometimes, it's difficult.
The pro even if you get a good result and you want to replicate this, you might need to replicate the experience and so on and so on. So that shows you that it's not sometimes, it's not really scalable.
So foundational models is an is foundational models are based on models trained on vast amount of data. Let me give you an example. ChatGPT. ChatGPT is the is the foundational model for for for text that probably we have all been using. You know? If you need to process a lot of data, text data, I don't mind you guys training your own model.
You are using a foundational model. Makes sense?
So if we're doing this with text, can we do that with time series data?
Yes. We can. And there are many foundational models for that. And the way to access them is with Connect. And the biggest difference is that training a model can be difficult, but using a foundational model requires mainly prompt engineering, which is much simpler.
And the the result would be something like this. And I told you, I'm from IT. I don't understand this.
So, Franco, would you mind to Certainly.
Yeah. So what you can see there is a simple example of where these models have been utilized, and this is the pressure for the discharge of a steam turbine. And for some of you might know steam turbines, we were able to fairly accurately with high accuracy predict what the steam turbine output would be using these models.
And there you can see the red line is the actual and the green line is the predicted.
You can see back in history how accurately it was predicting up to this point.
And yeah, the steam turbine discharge pressure is very important. It tells us how efficient this piece of equipment is operating. It can help the operator see what is coming and adjust parameters as needed. It can help the maintenance department or reliability department see when something is going wrong. Maybe the cooling system is needs cleaning or replacement. So it's a very useful tool especially if this is one of your key assets on your site.
Hopefully that covers it. Let's go to the next one. Oh yes.
And so there's something that's also valuable. You don't just get a train, but you can also do some weighted values to see what is influencing those trends. So the operator or the maintenance department can actually go and see what is causing that line to go up and down, and and pinpoint where the issues are and troubleshoot.
Thanks, Julia.
Great. So in conclusion, Connect enables you to use foundational models that can, for example, can be found the one that we we presented here is TimeGen that can be found in in platforms like Azure AI.
The third use case is about using AI agents.
AI agents is is a hot topic. Know? Again, probably will not it can mean different things from different people, but for this presentation, the definition that we will use is this software where you don't need to write the business rules into the platform. You don't program them. You allow the AI to take the decision how to complete the task requested by the user.
In this example that we're going to show you, we have enabled this AI agent with three tools. The first one, of course, Connect, in order to access the PI data. The second one is our document library where it is full with our procedures, forms, manuals, and finally, with the ability to send emails.
So let me show you this.
So this particular a a a agent can be integrated with different platforms. In this case, you are seeing with Telegram, but it could be with Teams or WhatsApp or whatever you want. So as you can see, I am asking, the agent to tell me the values of one of our assets.
So the agent is deciding what tool use. And in this case, it picked connect, and it was was able to provide the values of the asset that I asked.
Now after getting this information, I'm requesting from some x information about the I'm requesting for some documentation.
And as you can see, the agent this time didn't go to Connect. It went to a document library, and it's recommending me some documents. You see?
Now because I'm because I'm lazy, instead of copy and pasting, I'm just telling them, can you send me information to my email? And as you can see, this time the agent will decide to go to PI, to connect, to get the data. It will go to the email. It will compose the email, and it will be sent. And I hope it it will appear here.
It's a video, but I still get nervous that it's not gonna work.
It's there. You see? So can you see the beauty?
We didn't program these business rules. It was the agent who decided how to complete the task. But, again, how how would you see this in in our plans?
Yeah. Good question, Julio. I think we saw a similar example this morning in the session. You can imagine you're the operator on a chemical plant.
You've got a lot of processes you're looking at, a lot of trains, alarms going off. So your first focus is going to be operating plants stably and safely. Efficiency normally falls down the ladder there. So one of the things we were thinking of using this with is asking the AI to help you prioritize.
Where are you outside of the recommended ranges? What is the best things to focus on in terms of efficiency and reliability? And help you prioritize, and it can point you to the right procedures so you can follow them and go back to where you should be or how you should be optimizing the plant.
Yeah.
So in conclusion, conclusion, connect groups to be suitable to get connected with other AI agents. Like in this case, we're using n eight n.
So the last use case that we would like to share with you is about using Connect to train ad hoc machine learning models, where I think Franco will be the best to present.
Thanks, Julio.
Yes. So this last example is quite specific, but I can't share too much details. But hopefully, this will be useful to you.
We have a process on our plant that is manual firstly, so it's in the field making manual changes.
And it's nonlinear, which means if an operator makes a change, there's not a direct clear correlation to what that change is going to cause.
That means that these changes are normally time consuming and again they're left for the last time when night shift comes on or a weekend. So what we realized is it would help to train an AI. This is a perfect problem that an AI can solve, looking at history. So we used the traditional method to train an AI using Connect again, taking processed data and taking manual data that we've collected over time.
So what happens if you look at point number one there, that's really that first step, pulling in the process data and manual data, training the AI, and then the AI recommends with a visual output to the operator in the field what is the best steps to take next to optimize the process. The operator then makes those changes, feeds in the output of what is done into the model, it recycles that via connect to the AI model, it updates, it learns, and it recommends another change. So helping the operator to quickly optimize this plant without having too many iterations.
So this has been quite a good thing we've done on our site and it's helped us bring these the raw data that we've collected over such a long time and making it real for our operators, making their lives easier, making our plants more efficient.
And yeah, that's definitely one of the the optimal ways we've been able to implement AI.
Over to you, Julio.
So in conclusion in conclusion, guys, I believe that we are all in this room know how difficult it is to share data, but at the same time, how relevant it is. And from our experience, Connect has proved that, it helped us to simply significantly simplify the IT infrastructure, lowering the total cost ownership, you know, the TCO. But you know what's the most important?
It has allowed us to free the business to focus on the business problems rather than on the IT issues. Thank you.
