Real-time optimisation at the cement plant: Ioannis Sandis and Tejas Maru, CemAI (USA)

Okay. Ladies and gentlemen, thank you very much for being here. I would like to thank the organizers for the opportunity they gave us to present our company, CEM AI, our portfolio, and also to share with you some insights about the operational sustainability. We have two main products, as mentioned before. One is the predictive maintenance, and the other is the process optimization. The predictive maintenance solution is focused on reliability, on downtime cost avoidance, and on reduction to avoid also the loss of production. And the process optimization is focused on increasing the efficiency, reducing the energy consumption, and increase the throughput. CEM AI is a subsidiary of Titan.

It's an affiliate of Titan company. We started with Titan, and now we have installed our solutions around the globe in more than 20 plants and more than 40 assets, and we're continuously increasing our installation. And our team is a global team of cement industry experts from different backgrounds and spread all over the world. Both solution are working for the same target for the sustainability. Predictive maintenance achieve operational reliability, reducing maintenance costs, and also reducing the risk of production loss. With process optimization, we increase efficiency and also we minimize the consumption of energy, and we maximize the throughput.

We continuously try to improve our models. We cannot stop to say, "Now we are okay. We finished everything." No. We continuously upgrade our models. We're continuously integrating agentic AI. Our value proposition is simple and clear: to increase the margin of EBITDA, quick payback time, and also fast implementation. Now, these two solution, the process optimization and the predictive maintenance, are complementary, and both of them supports the sustainable cement production.

We have use cases for optimization, such as to decrease the clinker factor, very, very critical important sector, to increase the uses of alternative fuels, to reduce also the heat consumption, and to reduce the electrical consumption. At the same time, with the predictive maintenance solution, we try to preserve this sustainability that came from production. For example, we want to reduce the clinker factor. So we have two paths. One is the clinker, the kiln optimization to increase the alite, and the other path is the cement grinding optimization in order to increase the blyme, but keeping the balances.

In this process step, the process optimization identifies the optimum operating conditions, meanwhile, the predictive maintenance preserves these sustainability gains by reassuring that the equipment works as per their design specifications, that they work close to their intended performance. So let's say that these two solution creates a virtual cycle. Process optimization drives sustainability, predictive maintenance protects sustainability. Now, we will deep dive a little bit to the first solution, which is the predictive maintenance. And I would like to start by mentioning the basic methods of maintenance.

We have the reactive maintenance, the worst case of maintenance, which is practically troubleshooting. Preventive maintenance, the regular maintenance for the plant. And we have the two advanced methods, the predictive and prescriptive maintenance. Both of them are data-driven, and predictive maintenance also is based also partially in condition monitoring system, but let's say this is the old fashioned predictive maintenance. Now, both of them, they are fully based on data-driven, and what is the difference? Predictive maintenance is a part of prescriptive maintenance. Predictive maintenance predicts an eventual upcoming failure, meanwhile, prescriptive maintenance does something more.

It gives some kind of root cause analysis and provides suggestions for actions to be taken on site. The message is clear. The earlier we detect a failure, the easier and the cheaper the repair is, and also we minimize the risk of production loss. If we see the graph, our focus is on the left part, where actually we transfer the detection into prediction. What is the philosophy behind that? Our solution uses advanced unsupervised machine learning models that combines historical data, online data, and then detects abnormality much early in advance. And here is the magic. They detect the abnormality, even the abnormality of a signal, even if that signal is within acceptable operating limits.

A conventional automation DC system cannot catch this abnormality, but the machine learning solution gives the possibility to detect the abnormality early in advance. Our solution covers a plant end to end, from the crusher until the cement, from upstream to downstream. And it is important to mention that it is sensor-agnostic. It means that it can produce, creates value since the first day of deployment without the need of additional sensors. We need very simple information to start operating the solution, to deploy and operate the solution. We need historical data, online data, signal list, equipment list from the plant, and some mimics as taken from the DC system.

Here, I would like to say a little bit more because it's very important. In industry, the key success factor for an industrial reliability intelligence solution is the intelligence architecture. LLMs, models, et cetera, we can find, we can use everywhere, and now it starts becoming very common. But the architecture is the key success factor. And now, here you can see the most important layer of our architecture, which is the domain model. Domain model contains the process relations and the equipment hierarchies of a plant.

In spite that this is a predictive maintenance solution, the detection is not only coming from the prediction, I would say, is not only coming from the detection of a single signal abnormality, but also of an abnormality of a process relation between two signals, either from the same equipment or from different equipment. AI responsibility. In the industry, AI responsibility cannot be negotiable because human safety, equipment protection, and production loss are at stake. For that reason, we use a hybrid approach, which is a combination of advanced agentic AI and human-in-the-loop. There is a continuous communication between people from the plant and the dedicated service center team.

They daily analyze the events, they take actions on the plant. So our solution becomes AI responsible, a very important factor to drive sustainability because it helps integrating all the AI capabilities into the daily workflows of the plant and in the decision-making process. The user interface is very important to leverage the customer experience. It shows us the alerts, categorize them by area, by type. It gives the possibility of effective communication between plant and the service center team.

It visualizes very critical KPIs, and very important, it gives a very strong knowledge base for advanced agentic AI because it stores all historical actions taken of problems that happened in the past and could be similar with the problems that eventually happen now. Here, we would like to see an existing use case that refers to a kiln, and more precisely to the kiln roller. If you see... Oh, sorry. If you see there are here three signals. The red, which is the bearing temperature, the purple, which is the shaft temperature, and the green, which is the oil cooler temperature of the tire. And if you see the dotted, where is it? Pointer, here, yes.

If you see the dotted area, practically, with the first look, there is nothing wrong going there. But if we look closer, you see the green line from bottom goes to up. So what do we have here? We have a violation of the process relation between these three signals. Those signals are still within acceptable operating limits, but the process is violated, and now we have an event, an alert. And this triggered an action. The plant people took action. They fixed the leakage of the oil cooler, and then they measured themselves some benefits, 250K from the-- Those are the money that they saved for repairing or replacing equipment, and also they measured that they saved four days of production loss.

One more use case. This is related to process, and this is the blade deviation. We see an abnormality on the blade deviation, still within acceptable operating limits. But those triggered an investigation inside, and personnel of the plant found out that there was a problem in the clinker percentage of the mixing. And that was causing some flowability issues. They took action, reduced the clinker percentage, and then the operation went back to normal. Finally, I would like to stay here on a very important point. Those two solutions, the predictive maintenance and process optimization, as we said, they are complementary. And here we see something more than that.

We see the strongest points where these solutions practically make miracles together, act like catalysts, one to the other. If you see on the right side, this is the real-time optimizer, where actually, what does, Tejas will explain better later on. It gives suggestions to PIDs in order to optimize the process. But very frequently, happens what? Some PIDs may start being abnormal, unstable. These abnormalities, because those are process abnormalities, can be detected by the predictive maintenance solution. And further, there will be some suggesting actions to be taken on site, either to check the sensor or the actuator, or to optimize the PID parameters.

Then the predictive maintenance solution visualizes these problems. And finally, if there is a problem with the PID parameters, we activate our PID optimizer. And then the operation comes back to normal. That's from me. Thank you very much. And now I will give the floor to Tejas Maru. Thanks, Ioannis. Thank you. All right. Thank you, Ioannis. I guess we are going to run it as quick as we can because Peter is always stressed in the timeline already. So I'm going to run you through the optimization process, through using AI, very quickly, and if you come across any questions, you have a beautifully designed notepad on your table. Use it, and then we can discuss it later on.

So we'll start with the pain points that what brings AI optimizer into the process. And if you look at couple ones in the beginning, variability in the raw material, environmental restrictions, et cetera, the demand, the cost, it's been there ever since. But what I've been seeing lately that is a little new, is the loss of experience in the control room. Those 20, 30 years of experience that you had has been depleting so fast, and that's bringing a need of having this AI-based optimizer, which can then help stabilizing the plant operation by doing the job correctly.

So looking at the technology real quick, if you see, basically it takes the data from the plant, the machines, into the process database, and goes through a smart filtering process that eventually supplies the data to the control strategy layer that is built for having different scenarios for each asset. And from there, once it finalizes the process scenario, then it supplies the data into the AI-based algorithms, which then helps computing the set points for your process, whether the kiln or mill, whatever.

And basically, the takeaway here is that it comes with advanced analytics and smart data filtering that uses the AI-based algorithms to compute the process set points and outputs to help achieve your objectives. One other thing I have been hearing a lot, is garbage in, garbage out thing, when it comes to data, and sometimes, it appears to be a little more scary than how it is, because right now, the technology has evolved to a level where there is a workaround for this.

So this statement is not purely correct, and there is a way to deal with this, and that is where these AI-based soft sensors come into the picture, that would supply the soft data when your process is not having data because of any disruption or temporarily missing any set of data. Then soft sensors can supply data to keep the ball rolling. It keeps stabilizing the process until the sensors start working back. So this is something to keep in mind. Now, talking about the traditional old advanced process control versus the AI optimizer.

Back then, APC had a bit of a dependency where it was relying a lot on the targets that are given by the operators, and then sometimes restricts the performance of the optimized process, as it kind of becomes a victim of a mind of operator and where he wants to operate. Whereas these new AI-based optimizers are kind of pretty open, and all it requires is the final goals that you want to achieve and some of the hard process constraint that you don't want your process to go into.

And in between that, it kind of define its own operating level, which kind of gives it the freedom to optimize and operate at a level where sometimes operators are scared to operate the process or get rid of their mindsets. Talking about the UI, this can have its own independent UI, or the UI can be built within your plant SCADA system, and that way you don't have to deal with different screens. So it really depends on what the requirement and what the demand is in the CCR. Talking about the performance indicators, I normally call this slide blah because it's always 3% to 5%.

Everybody talks about it, but the reality is, and that's a fact, I believe that the real results that you get is always going to be 30% or 40% higher than what you see in here, because this data that you see is derived during the sense of awareness where operators know that he's competing with some kind of a system. But that performance is not going to sustain throughout the year. So, as I said in the first slide that with the loss of experience, with lots of diversions that you see, these systems have improved, and this has been performing a lot better than what you see in the results. Talking about a little case study here.

You see here a traditional preheater kiln cooler system, and this is a schematic of a process and a control diagram. So, handling the kiln process where it takes care of burning, traditional control parameters like burning zone, NOx, oxygen, CO, and it has a bunch of set points that it operates using the kiln feed rate and primary fuel, kiln secondary fuel, preheater fan, the cooler fans, as well as the cooler grate speed. And comes with three optimized targets with the priority of emission control, quality, maximize production. And each client, again, can change these priorities based on their requirement.

If you are in need of emission only or quality or production, you can always set the priorities on your own. And just quickly going through a case study where we came across a plant that was suffering because of the changes in the quality of alternative fuel. There was a high variation, and of course, the raw mix was also a little disturbed, where it was demanding a frequent change in the oxygen levels whenever the changes in the variation takes place.

And basically, to handle that, after installing the optimizer and by adapting these oxygen levels to handling the changes that it was getting from the alternative fuel as well as the raw mix, it easily helped the stabilizing in process by at least giving a good throughput of eight to nine tons, 30 to 40 kilocalories of energy savings. And if you look at the carbon reduction, that was almost around 3,000 tons a year, and that equates to a lot of value. All right, so two more slides, and I think we'll be done. The standard installation time is 8 to 12 weeks, depending on the plant and the requirements and what you have available, what challenges.

And the last slide I would like to talk is, are we ready for AI? Is another thing that I hear a lot. And we talk about data integration, continuity of special sensors, different technology trials, et cetera. But from the experience I've seen by doing this thing for over 20 years and over 200 plus installations, there aren't processes in the cement plant that would restrict you to install optimizers by any of these challenges. Of course, what you're going to see is the difference in the results that you would get. But I think the cement industry has enough amount of instrumentation ever since that always has a room for improvement by using this kind of system.

So what we normally say is that if your operators can see it on the screen, we can do it. You don't really need anything extra. Again, the results will vary, but it's always going to be plus. And again, summarizing both the solutions, you can see that there is no way you can go wrong either way. There's always going to be a benefit one way or other. It really goes down a lot to the culture. You want to adapt in order to have the systems operational and successful. But besides that, we have seen that it always works. Again, questions, please feel free to go over to the booth. We are right in front of the gates here. And thank you very much.