The Community for Technology Leaders
RSS Icon
Subscribe
Issue No.01 - February (1996 vol.11)
pp: 18-21
Published by the IEEE Computer Society
Steel production involves a number of stages, such as melting, casting, rolling, and forging, that entail complex chemical and thermic reactions as well as intricate mechanical operations. Because these processes do not lend themselves to exact mathematical modeling, steel manufacturers must turn to techniques for reasoning with incomplete and uncertain data. Their decisions often rely on the experience of individual experts. Nearly all steelmakers worldwide now use expert systems, fuzzy logic, and neural nets to improve quality assurance and production efficiency. This special track of IEEE Expert looks at several typical, successfully fielded systems.
For many years, the steel industry's main objective has been to maximize production by automating processes and streamlining plant organization. As with the Republic of Korea's Kwangyang Works, steel manufacturers have been erecting new plants from scratch, locating them near the sea to make the delivery of steel from blast furnace to final shipment as direct as possible. Because they restricted the diversity of their products, such new plants have become very competitive. Asia's steel industry, in particular, used these approaches in the 1980s to produce high-quality steel cheaper than its Western competitors. (See the sidebar for historical overview of steel production.)
However, the continuing improvement of substitutes for steel has raised the demand for even higher-quality steel with dedicated characteristics. By using different alloying metals and various heat and surface treatments, steelmakers now can offer a manifold of products. Ongoing research into new steel qualities has produced a broad range of products, which present many new control problems. Although other industries reflect the same tendency toward processing in smaller lot sizes, the steelmaking environment shows more diversity than most because of the particular characteristics of its material and manufacturing technology. Furthermore, the capital-intensive nature of the industry can make unanticipated violations of technological constraints extremely costly. A look at several typical factors will illustrate these considerations.
Most steelmaking processes are temperature-sensitive. For each process, the steel must have a prescribed temperature, and any time it spends waiting on the next processing step will incur a costly reheating. Moreover, because chemical reactions depend on temperature, any loss of heat during processing may degrade the steel's quality. If the prescribed solidification temperature profile is violated, an incorrect internal structure of the steel might result.
Although process times are difficult to predict precisely, steelmakers do exercise some degree of control. Treatment time in furnaces depends on the temperature, which can be controlled through heat input, generally subject to some global energy constraint. Steelmakers can also control casting and rolling times to some extent by varying the speed at which they run the caster or rolling mill. By slowing a process down if it appears that an order will arrive too early at the next aggregate, or speeding it up if it appears it will arrive too late, they can also use real-time process control to meet requirements for synchronicity. (An aggregate is a common steelmaking term for machine.)
Technological considerations in the production of higher grades of steel impose requirements on the sequence in which orders are produced. Chemicals added to steel to achieve certain characteristics react with the steelmaking aggregate. Residuals remain in the aggregate and may be absorbed by one of the next orders, which may be corrupted by this infiltration. The width and thickness of the steel product also constrain sequences in the casting and rolling processes. Production-run engineers will avoid some obviously incompatible sequences, but sometimes schedule incompatible ones anyway for lack of a closed tractable causal model that would prevent them from doing so. Afterward, causal models can explain these errors, and this negative experience will lead to a modification of the production process.
Automating production
The nature of these problems that complicate production control—the vagueness and always-changing nature of the knowledge—has prevented steelmakers from making closed control loops for steel production. This industry has always been very innovative in the application of new production technologies and the latest computer technology. It was one of the first, for example, to apply fault-tolerant computer systems to fulfill the high requirements for availability and reliability of control systems. Despite this rapid automation, the process operator has remained an important link in the production process, and with the introduction of these new control systems, operators are being overloaded with process data. (See the modern steel production sidebar.)
The steel industry adopted expert systems relatively early for further automating production. The five leading Japanese steel companies reported the first successful applications. 1 The designers of the Scheplan scheduling system, for instance, claimed that its operation saved $1 million a year by reducing the time that ladles carrying hot steel to the casters must wait. 2
Although quality optimization and energy consumption are important aims, the most important motivation for applying expert systems seems to be production standardization. In the steelmaking plant, for example, it is more important to have a safe and continuous supply from the blast furnace than a high, but irregular, supply. Because there is so much freedom in production decisions, it is better for quality assurance purposes to have formal rules about how to proceed, even if they are less than optimal. Having an expert system that acts as a consultant or even as a decision maker will make decisions more transparent.
Applying expert systems
Steelmakers apply expert systems instead of conventional software because the controlling software has to reason with existing uncertainties and master the inherent complexity of typical control problems. The control of the blast furnace illustrates issues motivating the implementation of expert systems.
The main focus of the blast furnace quality improvement effort is hot-metal silicon variability. This is controlled by the heat levels inside the furnace. If the furnace is too hot, the silicon will increase; too cool, silicon levels will decrease. Unfortunately, it is impossible to measure the temperature inside the furnace, so hundreds of sensors on the walls indirectly measure the hot metal's temperature.
Existing expert systems address two problems:
  • predicting abnormal situations such as slips (abnormal and sudden descendings of the raw materials charged in the furnace) and channeling (the heated gas reaches the top of the furnace without reaction) and
  • keeping the thermal condition stable.
Operators can adjust furnace heat levels and hot-metal temperatures by manipulating such variables as ore-to-coke ratios, blast temperatures, fuel-injection levels, and blast moisture levels. If the hot-metal silicon falls below the desired value, the hot-metal temperature will also be below its goal. The operator will need to increase the heat level. The problems in controlling this process are the long reaction times and the different reactions of human operators. Individual operators use different actions, start them at different times, and apply different magnitudes of changes. They also frequently use old data from previous cases to decide reactions.
Most blast furnaces managers therefore aim not to optimize but to standardize this control. The uncertainty of many data values makes it difficult to find a simple control algorithm. An expert system lets manufacturers build a model of the physical and chemical process in the furnace with symbolic values, abstracting from the thousands of measured data values. Rules allow the specification of certain standards—when and how an operator should react.
One of the first well-described systems in this domain was Nippon Steel's Artificial and Logical Intelligence System (ALIS), which controls several blast furnaces. 3 Comparisons between human and expert system performance showed that in 25% of the cases studied, the expert system performed better and only in 7% did the human excel. Furthermore, the system is continually modified to improve its competence.
This issue
Today, as the latest conferences on production control in steelmaking show, almost every steelmaker—from developing countries to the traditional steelmakers—applies expert systems. 4 One great challenge now facing the steel industry is to improve the self-adapting capabilities of expert systems. As mentioned, modifications of the production process are quite regular, and the tendency is to adopt even more flexibility. Intelligent steelmaking aggregates that adapt themselves to new steel compositions and requirements for the produced good are important research topics now. Research into intelligent organizers that can learn new strategies if the manufacturing objective or the production technology changes is also ongoing. The first article, by Nicolas Pican, Frédéric Alexandre, and Patrick Bresson, addresses this issue. They have developed a system that incorporates an artificial neural network to preset the parameters of a steel temper mill. They also show that a combination of AI and conventional techniques often solves industrial problems best.
An intermediate step in self-learning systems are systems that assist human experts. Because users of expert systems are not familiar with expert system techniques, they need simple techniques to adapt a system to new production facilities and strategies. The article by Jürgen Dorn and Reza Shams describes experience along these lines and makes propositions for improving this capability.
A focus of future research is the cooperation between expert systems in the steel industry. At the moment, expert systems are single-user, front-end computer systems dedicated to one function. Their integration with the existing organization is very simplistic. Most systems couple to a process computer or a production-planning system to obtain required input data. However, stronger coupling would increase the benefits of expert systems. The simplest solution is an expert system that performs this cooperation as its main task. However, a more generic approach would let expert systems cooperate in an open framework.
For example, a steelmaking shop scheduling system should receive knowledge of the status of the blast furnace, because the supply situation will influence the scheduling strategy. More useful would be cooperation between a scheduling system and an intelligent machine such as a caster that can decide which sequences are good and when maintenance operations should occur. Negotiations are also necessary between a steelmaking plant and its customers—the rolling mills and other plants. Because these plants operate under different sequencing criteria, a best sequence for one plant is not necessarily good for the other.
Despite even more pervasive automation in the future, human experts will remain unavoidable for production control in the steel industry, because new production failures that cannot be handled adequately by a system occur quite regularly.

References

Jürgen Dorn is a senior researcher at the Christian Doppler Laboratory for Expert Systems in Vienna, a basic research laboratory established by the Austrian steel industry to improve technology transfer from universities to steel industry. He received his MS and PhD in computer science from the Technical University of Berlin. He was involved in the development of two scheduling expert systems for the Austrian steel industry and works as a consultant for the international steel industry in the field of expert systems. He is member of the AAAI. Readers can contact him at the Christian Doppler Laboratory for Expert Systems, Vienna Univ. of Technology, Paniglgasse 16, A-1040 Vienna, Austria; dorn@dbai.tuwien.ac.at.
69 ms
(Ver 2.0)

Marketing Automation Platform Marketing Automation Tool