Dirk Schmidt & Eugen Geibel
KIMA Process Control
Contact : www.FTR.com.tr
Attaining ‘Industry 4.0’ has been seen as an essential task of the cement industry for many years, with terms such as ‘big data’ and artificial Intelligence (AI) heavily used. It is hoped that AI, combined with big data, will provide solutions to long-standing problems. It might therefore be surprising to learn that fully autonomous mill operation (including the use of AI) has been taking place since 2009. This article will briefly summarise how the methods of High-Level Control (HLC) were used in the cement industry already in the early 2000s and how they manage to control ever more complex closed-loop-controlled processes in situations where standard controllers fail.
To accelerate the integration of advanced tech- nologies in the cement industry, some business consultants proposed to ‘copy-paste’ Industry 4.0 solutions from chemical plants/refineries and apply them to cement plants. A recent example is a report regarding the first successful conversion of regular plant control to AI control, which was qualified as a breakthrough.
Caution: The capabilities of AI are still limited. This is in part because AI is a very broad term and it is difficult to find a definition on which everyone agrees. In the wider sense, it can be defined as a branch of computer science dealing with the simula- tion of intelligent human behaviours in computers. Technically speaking, the majority of the AIsys- tems that are used in industry today are data-driven algorithms. The basic principle is relatively simple, but they gain their capabilities from huge amounts of data, rapid and repetitive calculations and multi- ple interconnections.
The use of AI is not new. The development of faster computers with the possibility to store and process big data make the use of AI both possible and reasonable. Deep Learning, which is itself a part of Machine Learning, makes use of multi-layered artificial neural network (ANN) to learn from Big Data and search for patterns that could be used for problem solving after ANN training. Knowledge based automation, including fuzzy logic and analyt- ical methods such as, for example, Model Predictive Control (MPC) are also part of AI.
Figure 1: High-level control relies on AI.
Depending on the task, different AI methods are more applicable than others. Nowadays it is clear that there is no one-size-fits-all AI module for the cement production process. Fuzzy logic is applica- ble for closedloop control of technical processes with a moderate number of variables and data for which a control strategy can be expressed. It is a good choice for processes, where safe operation in critical situations is mandatory. ANNs are used for recognition of hidden process patterns for which a control strategy cannot be expressed and that have a high numbers of input variables. MPC is a good choice for well-understood processes for which a mathematical model is available. Optimisation is possible if the model can be calculated faster than real time. On-line adaptivity is definitely not practicable in cement production.
The AI in cement story
A short discussion with AI-solution providers makes it clear that the current ‘revolution’ in cement plant control sys- tems is simply operating with MPC and soft sensors. Germany’s Powitec Intelligent Technologies used self-adapting MPC and machine learning as early as 2001. It was the first comprehensive black-box controller that operated a rotary kiln fully autono- mously for more than 24hr without manual interaction. The core of the system was an image-processing camera that analysed the main burner flame and an online prediction of free lime. Various AI components were used as far back as 2002. Using adaptive MPC, the fluc- tuating energy input to the kiln and calciner was adjusted automatically.
Pioneers in the use of AI during these days were the cement producers Leube (Austria) and Maerker Zement (Germany). Soon after, companies such as ABB, FLSmidth, Pavilion, KIMA Echtzeitsysteme and Rockwell entered the market with similar model-based controllers. Today, many companies have returned to the more robust fuzzy logic con- trol. Predictive models are used for soft sensors. The reason for this are the serious differences in the cement manufacturing process compared to other common production processes.
Figure 2: The Loesche type LM 46.2+2 S VRM at the EQIOM Ciments plant in Dunkerque, France. See Case-study 1, overleaf.
Special process conditions
Clinker production is a complex operation. The so-called ‘multidimensionalnon-linear process model’ of a kiln or a mill has failed to model the real systems adequately. The real kiln or mill unit is subject to wear and other natural variations, which models have so far failed to represent. Broadly speaking, the behaviour of a kiln or mill tomorrow is different from its behaviour today.
There have been attempts at implementing software features, such as self-adaption and self- learning, but a host of changes has to be considered: liners, balls, chutes, feeders, valves, refractory, fuels and the raw materials themselves. And then there are even further changes to consider: the quarry components and additives, the fuel calorific value, water and ash changes, the change of coal and petcoke particle size distribution from their mills and the related change of combustion (ignition point, burn out, flame shape, etc.). All of these changes can affect the quality of clinker and cement – a significant challenge for any controller.
If a multi-dimensional process model is fed with signals that have a drift, or are not stable, it fails. It is possible to teach these models with an ANN in order to catch the drift and to adjust the model accordingly. The question is: How fast does the process change? If the changes are too rapid, the ANN based model has to be re-trained regularly, which is impractical for many manufacturing pro- cesses. It follows that for ‘drifting’ systems, MPC, is not the best solution. In such cases we need to in- clude some rule-based system to control the process when MPC or/and ANN results are not realistic.
Figure 3: An optimumly- filled ball mill chamber as a result of precise fill-level measurement reduces wear of liners and balls and boosts production.
Assisted by modern modules of AI, these sys- tems can reach a new level in automation. Below we detail some plants that operate their mills fully autonomously over several days. The ‘auto-pilot’ is not limited to smooth operation conditions only. KIMA’s MILLMASTER allows for fully automated start and stop of the mill, automatic recovery after emergencies and switching between cement types. We will also discuss case-studies where plants have increased their performance using KIMA’s SMARTCONTROL system. This software-platform has been supplied globally in nearly 200 rule-based MILLMASTER systems.
High-level control as a combination of different AI modules In 2008, KIMA Echtzeitsysteme (the previous name of KIMA Process Control) published an article about a project to supply 30 SMARTCONTROL packages for ball mills (including the SMARTFILL fill-level measurement system) to a selection of Holcim group plants in Eastern and Central Europe. After the commissioning and later, during steady operation of these plants, the development of MILLMASTER continued separately in the Holcim group as well as in KIMA Process Control. New designs of the human machine interface, pro- gramming logic and new software modules were developed to follow new trends in automation.
Figure 2 shows a combined grinding circuit, consisting of a roller press and a ball mill. Such a grinding circuit for cement is quite widespread because of its advantageous specific electrical energy demand and high product quality. From the view point of control theory however, this is a difficult system, with three sub-systems (roller press, ball mill chamber 1, ball mill chamber 2), each of which is critical. Each is also part of a respective feedback loop with individual delay times. The feedback loop enables each of these subsystems to oscillate on their respective resonance frequency, which is determined by the aforementioned time constant. Even worse, these time constants are non-linear functions of clinker quality (grindability), which is never guaranteed to be constant. Conventional PID-controllers are not able to handle such systems due to the oscillations.
Figure 4: Schematic of a typical grinding circuit, consisting of a roller press and a ball mill.
To better understand the complexity of the combined grinding circuit, such a system can be compared to three pendulums, which are coupled by springs with different stiffness levels, as shown in Figure 3. The time constant of each pendulum is determined by its mass and the length, the coupling is determined by the stiffness of the connecting springs. In the ideal case, this system is excited (shifted) by a constant raw feed, each pendulum moves to a new equilibrium state and remains there.
In practice, however, the excitation by raw feed is not constant and the length and stiffness of each ‘pendulum’ change with time. The result is a system, which is oscillating permanently at varying frequencies and amplitudes (Figure 6).
The task of a closed loop control system is to adjust the excitation (i.e. raw feed), the resonance frequencies (length of pendulum, i.e. transport speed of bucket elevator, conveyor belts and air slides) and the stiffness of the coupling springs (i.e. mass flow from roller press into chamber 1 and from chamber 1 into chamber 2). Controlling such systems is a difficult task and cannot be performed by using a single PID controller. According to KIMA’s experience, such a system can be controlled successfully with MILLMASTER.
Figure 5: Mass-spring model of the combined grinding circle with a roller press and a two-chamber ball mill.
Figure 6: Excited spring-mass system.
Case study 1: Fully autonomous mill operation at EQIOM Ciments
At its plant in Dunkerque, EQIOM Ciment’s Loesche type LM 46.2+2 S vertical roller mill mainly produces slag cement. It is controlled by a MILLMASTER system. From Friday after- noon to Monday morning the plant is operated completely autonomously under the control of a MILLMASTER system.
“The MILLMASTER system is used daily and gives us the opportunity to concentrate on perfor- mance optimisation while the mill is running. It is also faster than an operator when it comes to pro- tecting the equipment in case of important changes in mill behaviour. It would be hard to run without this expert system for a long period,” says Pierre Vonstein, Operations Manager for North and Normandy Grinding Stations at EQIOM.
Case study 2: Increased production at Fabrika Cementa Lukavac
A key advantage of the MILLMASTER system is that it can be configured in such a way, that the operators do not see much from the system. Just the ‘on/off’ switch allows them to start or stop the ‘auto-pilot’. Switching it on is only for one purpose: increased production. A representative example is Fabrika Cementa Lukavac in Bosnia & Herzego- vina. In 2018, the system was installed on a ball mill rated at 65t/hr. The mill usually reached a base line of 67t/hr, according to Process Manager Emir Cilimkovic. Table 1 shows the before and after situa- tion at the plant. There was also reduced fluctuation in quality and less wear on the grinding media due to a higher fill level in the mill. The optimal filling level was confirmed by means of the crash stop method.
“The MILLMASTER is a useful tool for us to achieve lower standard deviation of cement fine- ness, which worked out to our full satisfaction,” reported Cilimkovic. “Today, we need 2hr less to produce the same amount of cement as before. Power consumption has been reduced by 9-11%. SMARTFILL also reliably helps to detect blockages in the intermediate diaphragm of the mill. This feature has become especially important because of the increased usage of alternative fuels in cement kilns and the ongoing change of the clinker chem- istry. For maximum production levels, we carried out extensive checks on the system from time to time. The filling level inside the chambers and the condition of our balls and liners is checked using the crash-stop method.”
Case Study 3: Increased production at Opterra Karsdorf, Germany
After the merger of Lafarge and Holcim, plenty of plants were told to switch off their former expert systems LUCIE (Lafarge) and MILLMASTER (Hol- cim), as they left the group. Both expert systems required a certain level of support by technology centres, and experts visited on a regular to secure successful day-to-day operation. Following this, ‘a couple of well-known suppliers’ were asked to equip these plants with an alternative software that offered the same performance as the previous systems, but, if possible, with easier handling from the plant site and without the necessity for regular maintenance from external resources.
Figure 7:View of the Cementa Lukavac plant in Bosnia & Herzegovina.
Table 1: Production increase at Cementa Lukavac after installation of the MILLMASTER system.
In 2015, KIMA Echtzeitsysteme was awarded a contract to equip all ball mills at the plants that had been acquired by CRH in Germany. While there were a few concerns about the small, albeit well established, supplier KIMA, it was also known that this company had previously equipped some 30 Holcim plants in Eastern Europe with SMART- CONTROL systems. The former Lafarge plant in Karsdorf was also equipped with the MILLMAS- TER expert system. A total of six cement mills, four of them centre discharge mills, received an individual software package and parametrisation. This was to reach challenging optimisation guaran- tees of either a 5-8% production increase or 4-6% reductions in specific energy demand. All of the performance goals of the project were achieved, with full acceptance by the operators.
Conclusion
Many believe that AI consists soley of self-learning ANN that feed on big data to automatically control complex processes. This is not accurate. As already mentioned, different AI methods should be chosen depending on the task. It seems that the complexity of the clinker burning process makes it a bad candi- date for pure ANN control, not least because of the wear of the involved equipment, which changes the system that the ANN was initially trained for. The significant components of AI being used today are similar to those used 10-20 years ago. However, to- day’s computers have become much faster and have access to large amounts of historic data.