As for SIBUR, we have the Digital Technologies function, one of the divisions engaged in advanced analytics. We participate in all processes that are somehow related to data analysis in the company: from data from a multitude of sensors at production sites to stock reports and forecasts. Using this data, we develop digital products that help companies significantly improve operational efficiency.

My name is Alexander Krot, I represent the advanced analytics division, and I will share with you the following:
 

• how to minimize unplanned downtime of an extruder that cuts polypropylene into granules and tends to get clogged
• how to use data analysis and a customized model to increase butadiene production by more than 100 tons per month
• why it is better to predict reactor problems, instead of examining the reactor using X-rays

Operating efficiency can be improved in two ways. The first is to increase production output through a more stable mode of production. The second involves a reduction in losses from unplanned outages, each of which leads to losses of millions of rubles. 

Very often, people think of a working model that will be used for predictive analytics as just a program. They believe that it is enough to give us a code that will help to predict possible problems. We decided to go the other way. We create ready-made models as comprehensive turnkey products. That is, there is always a working self-learning model that can be adjusted as well as convenient interfaces / monitoring terminals and support that respond to user requests for new features. Our users are operators who monitor the production process, so they know exactly what needs to be changed in the product to make it more convenient. After all, it is the operators who will monitor the terminal, respond to changes in variables and make adjustments.

Therefore, we create such models according to a classical product scheme, where a team is formed with the participation of a product owner, developer, designer, and data engineer for each particular product. In addition, there is always a process engineer in the team who understands the features of the production process that we intend to improve.

Each of our projects lasts from 3 to 6 months, depending on its complexity. To begin with, we conduct reconnaissance by sending a team to the production sites; we have a research procedure called "framing," which allows us to determine exactly what the customer wants and whether it is possible to solve the problem with the help of data. If the answer is yes, then we continue our assessments: It is important for us to understand whether there is enough data available to solve the problem, where the data can be downloaded from, and whether organizational changes will be required. Of course, we will separately calculate the economic effect in order to subsequently rank projects and implement only those that give the maximum return. It is clear that if we spend a lot of time and effort on a project that ultimately saves RUB 5,000 per month or a couple of light bulbs, then such a project is not advisable.

But if we understand that the product will bring real benefits to the сompany (both from the point of view of optimizing operations and improving working conditions for the staff, and from the point of view of financial benefits), then we start working. We have already implemented almost a dozen different projects using this approach, and in this article I want to highlight a couple of the most significant of them.

One of the goods that SIBUR sells is polypropylene, we sell it as granules in bags (more details on our products can be found here). The production of polypropylene from gas involves several stages, but here we will discuss the last one, which consists of cutting granules. There are peroxide types of polypropylene when peroxide is added to the homogenized mass of polypropylene. That is, peroxide is added to the molten flowable polypropylene, and this mix is then fed into the extruder.

An extruder is, in essence, a huge chopper. Its size is equal to an average two-room apartment. In this case, we pay attention only to certain parts of the extruder, namely, the screw (like a meat-cutting screw, this screw mixes the melt with peroxide), the spinneret (this is like the mesh in the meat grinder, to which a mixture is fed under pressure) and a blade cartridge that actively cutts polypropylene "pasta" into granules at the back of the spinneret. The cut granulate rises to a special vibrating screen with air flow; then the granules are packed in bags, and they are generally ready for transportation.

For various reasons, unplanned stops of the extruder occur.

For example, the peroxide has not been mixed well because of failure to comply with temperature regulations or insufficient pressure, etc. As a result, this mass begins to stick between the spinneret and the blade cartridge. Therefore, instead of normal granules, an agglomerate is formed, which rises up and clogs the vibrating screen.

If this agglomerate becomes visible, it means that everything in the extruder itself has already been clogged. It is necessary to stop production, disconnect and disassemble everything, remove parts, clean the spinneret and blades. This usually takes more than half an hour, and the company incurs serious losses.

And now you will see what data science has to do with this.

In 2017, there were 19 such shutdowns. We collected data on them and looked at the parameters of the process mode for extrusion and polymerization. The analysis helped to identify a number of patterns. The result was the creation of a model that notifies the operator when something might clog about an hour before such an event actually occurs.

We delivered a full-fledged system to operators. Now they have an interface and several screens that display all the telemetry associated with the production process in real time. In order to simplify it, we highlight the necessary data in different colors (green-yellow-red), like on a speedometer. Moreover, when looking at the extruder itself, nothing can be assumed; however, the system uses telemetry and instrument readings to generate alarm signals that in 2 hours (this is the forecast horizon of our model) the extruder may clog. In addition, the system will prompt the operator that, for example, if he/she presses the blades harder right now, an unscheduled stop can be avoided.

One of our main tasks is to ensure the survival rate of our tools. Operators should trust the system. If the system often generates false alarms, and the operator is has to recheck everything, then sooner or later he/she will stop responding properly. Moreover, the operator may get the impression that this is not a very clear system, which sometimes operates falsely and prevents them from working normally. Therefore, we additionally trained the model to minimize false positives. We installed video cameras over the vibrating screen so that the operator could observe the process, and if the system misses something important, the operator can visually observe the agglomerates beforehand and not when the entire vibrating screen is clogged. When employees change the blades or extruder settings, they immediately notify the support staff and ask to take this into account so that the model works more accurately.

Butadiene is our intermediate product, which is used, for example, for the production of synthetic rubber. Butadiene production has one important feature: it needs a fairly valuable catalyst. We usually buy it in quantities sufficient for 2–3 years in advance and pay several billion rubles. The catalyst is so expensive because it contains precious metals.

We have 2 reactor units with 8 reactors in each. Without going into details of the process mode, we can generally describe the operator's work as follows: you have a predefined temperature (we call it the "setpoint"), and it must be maintained during your shift. The temperature is controlled with the help of air dampers. The operator must maintain the temperature at the upper limit of the allowable range so that the catalyst does not burn and the yield of the final product is maximum. Essentially, the main task is to maintain the stationary mode as much as possible.

If you manage to keep the temperature close to the upper limit, the production will be sufficient, and there will be no harm for the catalyst. It would seem that just fixing the temperature is enough, but many different factors can have an impact.

It is worth stating that the work of the operator is also not so simple. Any action to change the temperature by opening the dampers has a residual effect over several hours. It's not as easy as in the shower at home where you turn on the hot water and realize that it it too hot, so you add cold water, and then everything is okay. But even in this case, the situation may change if a washing machine begins to use water, or if one of the neighbors also decides to take a shower.

As for our case, the following happens. You open the damper by 1 degree, and before you can evaluate the effect of this on the overall temperature change, you will need to wait for a while. In general, the operator turns the dampers back and forth on average three times during one shift.

We collected historical data and determined how much the temperature changes when the damper opens by 1 degree. By 2. By 3. As a result, we built a set of models that have become the de facto advice-giving system for operators. If the temperature is suddenly somewhere different from the setpoint, the system immediately generates an alarm signal and tells the operator which valve must be opened and by how many degrees in order to ensure the optimum temperature. The operator immediately sees this and responds properly.

This project is considered to be predictive maintenance. This approach helps us a lot. For example, we predict when and where something can go wrong, or when the oil or bearing needs changing, and we order the necessary parts in advance so that by the time of the event itself we only need to take and install them, instead of solving issues with ordering, delivering, etc.

Now, let's move on to optimizing the production by maintaining an optimal mode.

There was another effect from our model: the collective image of the operator has changed. Operators became more diligent and attentive, and compliance with the specified temperature range was included in their KPIs. Now they discuss which shift coped better with the task and quickly learn how to use the new functions of the model. In general, we gave them a good tool for work, and they give us quality feedback to improve this system.

At the end of the shift, the system automatically generates a report on the effectiveness of each operator, which makes it clear who has the right to boast. In addition, such approaches change the very work culture at the production facility. The very image of the operator has also changed — the role has become more digital, and now operators understand and use digital tools, possessing all the necessary skills. They are also actively involved in the development and improvement of these tools. 

Predictive model of coke formation in the butylene fraction looks like this. Meshes with thermocouples (thermal sensors) are installed in the reactors. During the reactor operation, coke often sticks to these meshes, which leads to their destruction and increases the repair time during shutdown. When this happens, the reactor is stopped and cleaned, and the damaged elements are replaced. Downtime takes about 7 days. The idea was to predict the formation of coke and burn it during a short shutdown without opening the reactor itself, which would have allowed an increase in the turnaround cycle.

How can we understand that coke is beginning to accumulate in the reactor? We can X-ray it. But this entails significant financial costs. Therefore, it was decided to optimize costs and use analytics.

When coke begins to stick to the temperature sensors, they usually show a slightly lower temperature, as well as a lower temperature dispersion. We observed this and built a model that started to predict coking without gamma scanning. This model is still at the pilot stage, but it has already provided the following features:
 

• A unified interface for monitoring all the sensors on all meshes
• Advance understanding and scheduling of repairs and repair personnel workload 
• Longer turnaround cycles and reduced downtime caused by repairs.

It may seem that analytics in production relates only to production itself. This is not actually true, and we also have marketing use cases. For example, we can predict market prices for certain product types.

It is important to note again that we do not build models for the sake of the models themselves; instead, we create finished products based on these models. Therefore, we have also created an ML framework, which has actually become the unified standard for model requirements. Regardless of which team created the product (it may even be a third-party contractor using the API), it is important for us that all these models use a single interface. This allows us to understand which models are working correctly, which models are beginning to degrade, which models are not working at all because of the lack of data, and so on.

When there were only 5 models, everything was simple in terms of both monitoring and support. But when there are more and more models (including products from contractors), an ML framework is a great aid, allowing each digital product to be stored — in a harmonized manner — in a container with subsequent automatic deployment of the API. We can put all the models in there and monitor them simultaneously.

This is why we use our own framework.