Pictures of the Future    Spring 2006

Formula for Efficiency

It’s been 18 years since Siemens first started developing learning software. Over that period, the company has assembled a library of methods that boost the efficiency of systems, ranging from rolling mills and power plants to genetic engineering, dishwashers and logistics.

Learning systems have suddenly become big business. "Internet companies such as Google and Yahoo are hiring lots of people who know how to handle these systems," says Dr. Volker Tresp. Together with coworkers from the Competence Center for Learning Systems at Siemens Corporate Technology (CT) in Munich, Tresp develops software solutions for steel mills, washing machines, and data mining—the analysis of large volumes of data—in other words, precisely the kind of technology that’s needed in order to enhance search engines. And yet this research group has been around a lot longer than any of the Internet giants. In fact, it was established 18 years ago. That made it one of the first industrial research groups in the world to take an interest in learning software. "But we’re just as much involved in the academic world as in industry," says Prof. Bernd Schärmann, Director of the Competence Center. Besides developing software for pretty much all of the Groups at Siemens and applying for a variety of patents, the 30 researchers in the Competence Center have also published around 200 papers in scientific journals.

The department was set up in 1988, back when the promise of so-called artificial neural networks was at its peak. Neural networks are computer systems that process data according to principles similar to those that govern neurons, or nerve cells, in the brain. In contrast to conventional systems, no fixed rules are programmed for neural networks, because they are able to learn from experience and adapt on their own. This occurs in the course of a special training phase involving different examples. If, for example, the network has to learn to identify cars on the road, then it can be trained to do so on the basis of images of different vehicles.

Machine Learning Library. In fact, the idea of neural networks is 50 years old, but it was not until the mid-1980s that their properties were extensively researched. Back then, even the first, primitive networks looked so promising that scientists were soon persuaded that they could be used for just about any application, ranging from pattern recognition to autonomous robots and even weather forecasts. Yet the euphoria soon died down with the recognition that neural networks often had only limited utility in practice. This was a result of, among other things, the lack of sufficient training patterns.

Today, researchers know that the networks tend to perform best when combined with other procedures, such as fuzzy logic, for example, which recognizes not only binary values but also intermediate ones. Statistical methods also help performance in that they generate predictions on the basis of probability theory. "All of this suites us fine because our strength lies in the breadth of methods that we have at our disposal," explains Volkmar Sterzing, who is head of the Advanced Control, Prognosis and Diagnostics team at Learning Systems. These methods are assembled in the so-called Siemens Machine Learning Library (SML), developed by engineer Bernhard Lang. SML has already played a key role in ten products and systems from five Siemens Groups.

One such product is SIMELT SIMPAX, a solution used in the direct reduction of iron ore to pig iron. Instead of being smelted, the iron ore is chemically reduced—that is, it’s deoxidized by a flow of hot natural gas, which strips off oxygen compounds. The process is difficult to control and demands a lot of experience, since parameters such as temperature and the removal of the iron have to be precisely set in order to ensure metal of a requisite purity. Moreover, the quality of the end product can be determined only at the end of the process, by which time it is too late to modify the parameters. Lang therefore developed a prognosis model based on learning networks and thermodynamic formulae that simulates the process in real time. Based on experience and examples, the network has learned which input values—concentration of added gases, temperature and throughput rates—result in the purest iron.

Improved Rolling Power. Although there is nothing new about the idea of process control, what sets these solutions apart is that they aren’t static but are adaptive and capable of learning. "The system itself finds out how much energy the facility requires, when the temperature has to be raised, and what the optimal pressure is, rather than having these values programmed in," says Sterzing about an application for power plants. "The software then automatically discovers the best configuration between the various parameters." Engineers at the Power Generation Group (PG) are full of praise for the applications developed by Sterzing and his colleagues. "This software makes us the technological leader," says Uwe Gerk at PG, who helped set up the learning software project.

The breakthrough for Learning Systems came in the early 1990s with a solution for rolling mills used in the steel industry. A vital parameter in this process is the hardness of the steel, since this determines the force with which the metal has to be rolled. Thanks to a neural network—first introduced in 1993 and continuously enhanced ever since—rolling power can now be gauged with 30 % greater accuracy than before. "It’s established itself as the industry standard," says Dr. Einar Broese from Siemens Industrial Solutions and Services.

The control software is now so advanced that it is based upon the precise physical characteristics of the rolling process so that the neural network can determine the yield stress on the basis of the chemical composition of the steel. What’s more, it learns on the job, modeling its future performance on the best results.

Neural networks are now also used to determine other physical parameters in rolling mills, for example to calculate the requisite temperature or to predict the width of the steel band produced by rolling. The latter is important, since customers demand a minimum width, and steel producers are naturally anxious to avoid any overhang, as this uses up more material. At a plant in Duisburg, Germany, it has been shown that the use of a neural network enhances width prognosis, thus leading to a reduction in material requirements of almost 10 %. At a current price of several hundred euros per metric ton of steel, this means annual savings of millions for the Duisburg plant, which produces around four million metric tons of steel a year. Neural networks from Siemens are now in operation at around 60 rolling mills.

Learning software can be used profitably to control almost any kind of industrial process, including those at power plants, rolling mills and paper mills. What’s more, it is also suitable for completely different applications, such as predicting the demand for a particular product (see Pictures of the Future, Fall 2003, Logistics). "Advances in modeling have now made it possible to predict complicated systems," explains Dr. Hans-Georg Zimmermann, whose developments include a model to forecast electricity prices. In the past, this was easy to predict. Power was cheap in the summer, and expensive in the winter. Today, however, the market has become fragmented, and companies have to buy at least a month’s worth of electricity in order to obtain a favorable price. Besides using the model itself, Siemens markets it via Siemens Business Services. "Learning systems allow us to predict every stage of the power-generation process. First, what price the power plant has to pay for coal or oil; second, the production conditions, for example the amount of pollutants produced by the power plant; and, third, the price at which the power can be sold," Zimmermann explains.

Focusing on Genes. Meanwhile, Siemens researchers are working on other applications, including the recognition of cancer genes by means of biochips that analyze a person’s genetic makeup. Interpreting the results is complicated when lots of genes are involved. Dr. Martin Stetter has developed a model that relates individual gene activity. In short, it can track down gene clusters and predict the probability with which a gene will be activated when two other genes are also active. This, in turn, enables faster detection of key genes. A pharmaceutical company has already tested the software in a pilot study (see Pictures of the Future, Spring 2005, Molecular Detectives).

Learning systems will soon be helping at home too. Sterzing’s team is now working on a contract for household appliance manufacturer Bosch and Siemens Hausgeräte GmbH to improve dishwasher performance. They’re developing a dishwasher that can gauge the amount of dirty dishes—like modern washing machines that automatically regulate water and detergent amounts in line with a load’s weight and how soiled it is. This is important, because a half-full machine uses less water and electricity than a full one. Sterzing won’t divulge precise details, of course, because rival manufacturers have also started to use learning software in this field. But the Learning Systems researchers have one major advantage: They’ve been in business for 18 years.
                                                                                                          Jeanne Rubner