# what the rolling mill 2000 lacked

We have a rolling Mill-2000. 2000 is not the year of creation, but the length of the working roll barrel in millimeters, which means the width of the steel strip. The mill rolls a metal ingot called a slab (about 25 tons, 25 cm thick and 10–12 m long) into a steel strip up to 1.5 km long.

Everything that enters the mill from one side will travel as much as 1,500 meters to the very coilers, will be rolled out and wound into a finished roll. At the output, the strip thickness is at least 1.45 mm, and on average about 3 mm. Since the mill is long, several slabs/strips run on it at once: while one strip at the end of the mill is wound into a roll, the other rolls, and a couple more slabs are already entering the mill from the furnace.

This is a key link in the entire production. The productivity of the shop is largely determined by how well we load the rolling mill.

Naturally, we want to load as many slabs as possible onto the mill, but at the same time just in time so that they do not interfere with each other. Therefore, it is important:

So that the slabs do not catch up with each other, because then there will be a traffic jam and everything will stop.

In order for the gap between the slabs to be minimal, that is, they must be fed almost end-to-end.

The challenge comes down to understanding the ideal moment to feed the next slab to the mill.

Before the advent of the IT system, which we want to talk about, people generally coped well, but they always played it safe and laid a few extra seconds of pause between slabs (such a pessimistic forecast instead of an optimal one).

After all, the complexity is also in the fact that, depending on what size of the strip you need to get, you need to roll it at different speeds. And there are also completely non-algorithmic factors such as “somehow the band is not shaking so much” and “something I don’t like the sound” – we still cannot single them out for full automation.

That is, we need a system that works in tandem with a person (more precisely, a team of four people): the model makes a decision, and people can make adjustments if their experience tells them to do otherwise.

This is a hot rolling section: before the metal enters the rolling mill, it is heated in a furnace and acquires the softness and plasticity necessary for processing.

This is the issuance of red-hot slabs from the furnace, i.e. the beginning of the rolling process:

This is one of the five furnaces in which slabs are prepared. They are “prepared”, but not “prepared”, because they were smelted earlier in another workshop, about which we already wrote, and here they need to either be heated after intermediate storage, or to keep the temperature in the desired range before they enter the rolls mill.

Furnaces are arranged in a row and operate in turn. It’s like a revolver drum, from where slabs, like bullets into the barrel, are fed to the mill. But unlike a revolver, we don’t have a drum here, but an elongated structure, so the time for each slab to approach the mill from different furnaces is different: it’s not far to go from the nearest furnace, 10–15 seconds longer from the far one, and the system takes into account this discrepancy for the correct rate of supply of slabs to the mill.

Here are the five stoves on the diagram. Each contains up to 40 slabs, and they have different heating curves depending on the chemistry of the metal. It is impossible to underexpose the slab in the furnace, but it is possible to overexpose it a little: this is energy consumption, but it is not as critical as not getting into temperature and delivering a cold slab to the mill. But from the point of view of our site, all furnaces are just a “black box” with the property “give a slab in N seconds”.

The measurement of pauses between the strips occurs in front of the sixth stand – this is the beginning of the finishing group of stands. By the way, there are 12 stands in the camp.

The pause should be ten seconds (this is the safety interval to cover various random processes in the mill). During normal operation of the mill, a long pause between the strips before the finishing stands is already a luxury: a decrease in productivity.

People, on the other hand, took longer pauses, because there is slippage of the strip on roller tables, there are incorrect sensor responses, there are forecast errors, tracking errors, and signal delays. There are many possible situations, but the solution is one: add time.

## How is tempo different from pause?

A good example is buses. Pace is the time between the arrival of one bus and the arrival of another. A pause is the time between the departure of one bus and the arrival of the next.

## How a slab becomes a strip

The slab (let me remind you: this is a 25-ton metal bar), passing through the first part of the mill (roughing group, the first five stands), rolls out to a thickness of up to thirty millimeters and becomes longer. Further, it is no longer a slab, but a strip, and it enters the second group – seven finishing stands. In each stand, the steel is compressed by rolls, due to which it becomes thinner and longer. This happens not only unpredictably, but with variability in the mass of parameters that affect the release time of one or another section of the mill.

Then the strip enters the cooling section, where it is washed with water from all sides, and in the final it is rolled up.

## Risk versus Performance

Each strip size is rolled for a different amount of time. Even in the same size, the machine time may differ.

The team controls the speed of rolling, but within certain ranges, fixed in the technological instructions. It is determined by the initial and final dimensions, reduction parameters, cooling work and the hardness / softness of the steel grade that we roll, whether all the stands are in operation (some can be excluded from the process for dumping and filling new rolls, they work without deformation for about three hours ), and a whole bunch more. The team analyzes everything and decides how to go.

Of course, the crew is interested in maximum productivity, but the more they increase the speed, the higher the risk of sticking the strip in the mill (full paragraph). And then it pops up that you go quieter – you will continue.

The slower, the fewer risks, of course, but performance also drops – this is how a double-edged sword and eternal tugging of opinions turns out.

## Operator and mill

Also, the rolling speed (actual, and not even the one we want to set) depends on whether the strip is shifting or not, how well each hydraulic pressure device works, and also, with what sound the strip enters the stands, which is generally happening now with the mill as a whole. how cheerful and self-confident the team is today, and so on. There is no “autopilot” in hot rolling, the responsibility lies with the operator.

In our mill-centric model of the world, the operator is an external neural network that, over several decades of mistakes, has learned to manage the mill quite well. At the same time, each team developed its own decision-making model. Everyone works out an absolutely identical situation in different ways, so the number of the brigade for us is also one of the parameters of the state of the mill: a very informative indicator.

But we are by no means trying to unify the brigades. The main task is not to dictate to the mill operators what parameters to set, but **try to predict** how long will it take for each lane to pass under the parameters set by the operators.

To predict, we have statistics on how different brigades rolled different grades of steel under different conditions of the mill.

Of course, on the basis of the current input, the decision of the brigade cannot be guessed, but it is necessary to approach it in time estimation.

## How it works now

At the entrance to the process, we have a daily rolling schedule. In the schedule – everything that will be manufactured.

We constantly recalculate the forecast: we issued one slab and immediately recalculated the forecast for all slabs remaining in the furnace, taking into account their standard size and other parameters that affect rolled products. We need to provide a pause of exactly 10 seconds between the exit of the strip from the sixth stand of the finishing group and the entry of the next strip into the sixth stand: this difference is the pause.

This is counter-intuitive, but slabs from the second and fifth furnaces can enter the roller table at the same time, however, due to different distances and processing times, they will arrive in turn with the required gap.

Next, we record data on the state of the mill and need to know what is in the furnaces: how many slabs and what temperature, how long it takes to heat up the slab, etc.

Then we take into account mill telemetry. In 10 minutes, about 100 Mb of “raw” data is collected from sensors throughout the mill. We do not need all this lot of data, but only the main reference points and aggregates on which we built the training sample for our “smart” service are needed.

We put them in storage for a long time and saved them up, and then analyzed them. From this data, we learn who rolled, what condition the equipment was in. This makes it possible to predict how much the current order will be rolled on the current state of the equipment. The model predicts the rolling time, adds a safety interval of ten seconds, and we get data to order a slab from the furnace.

Five LightGBM models are considered:

Forecast of the strip passage time in the roughing group of stands.

Model error prediction for the draft group.

Reverse forecast.

Forecast of the strip passage time in the finishing group of stands.

Model error prediction for the finishing group.

All this is considered on a regular server with a CPU, even a GPU server is not needed.

## Operator and service

Data on the pause between slabs goes to the process control system of the mill, the operator sees these values and can correct them: he can manually change any values, for example, increase the pace, change the pause, or go completely to manual control. Unlike many of our other services, this one is not advisory, but managerial: calculations from the service are transferred directly to equipment management in the event that operators do not make adjustments.

On average, the model error is 1.5 seconds. This means that the model is about as wrong as a person with 30 years of experience is wrong. But this is exactly *average*. The fact is that on popular assortments (80% of rolled ones), the model has an accuracy of 0.5-0.8 seconds, that is, it gives a significantly greater gain. But unpopular complex technical maps are the accumulation of data for subsequent retraining. In the draft group, the percentage of prediction is better, in the finishing group it is worse, there is room for improvement in the model.

At the beginning of the implementation, the operators, understandably, did not have much confidence in the values of the service. But we tried to explain and involve – gradually people began to help the service become more accurate: to report errors in time and participate in the analysis of their causes, and now they just use it.

The result of the results of several months of work is a total reduction in delays by five hours. That is, it is as if we worked five extra hours non-stop. We count according to the average assortment, that is, as if in the past similar periods we would have rolled the same steel. For comparison, periods with the same state of the camp are selected.