Where Experience Counts

Brilliant minds are behind every new achievement in artificial intelligence (AI). Members of the scientific community include Volkmar Sterzing and Steffen Udluft, who focus on trends in the industrial use of new AI methods at Siemens Corporate Technology (CT) in Munich. They have achieved prominence in professional circles with practical applications of reinforcement learning for controlling combustion processes in gas turbines. Their current work focuses on new applications for genetic programming. Both researchers were distinguished in 2017 with the “Inventor of the Year” award from Siemens.

“Things always get interesting for us when a complicated system has to be controlled because so many measurement and control variables come into play,” says Sterzing, who leads the research team Learning Systems. This is the case with the combustion process in a gas turbine: The cocktail of gas and air travels from the valves into the chamber, where it ignites. Temperatures climb to levels of more than 1,600 degrees Celsius. How the gas burns, whether powerful combustion dynamic occurs, how much nitrogen oxide is created and how long a gas turbine can ultimately stay in operation – these are all determined by a number of factors: The quality of the gas plays a role as well, just as the outside temperature and the required combustion performance. For this reason, many experts are working to determine how the combustion process can be optimally controlled. This need prompted Volkmar Sterzing and Steffen Udluft to come up with the idea of using artificial intelligence to control a gas turbine.

When Artificial Intelligence learns with Little Data


From the outset, they committed to using methods that need relatively little data for the learning phase and, in doing so, they broke new ground. Accomplishments made by other AI researchers – for instance, with a software that defeated world champions in board games such as Go and chess – were all founded on the use of immense quantities of data. The gas turbine would need to run about one hundred years to produce a similar quantity of data for the turbine's training program. Faced with this problem, Udluft developed a data-efficient method for reinforcement learning with neural networks. This put the team from Munich ahead of the pack worldwide.

No one will buy ‘dumb’ devices and systems anymore when smart ones are available at comparable prices.

Today, Siemens uses a system called GT-ACO (Gas Turbine Autonomous Control Optimizer) in pilot operations to control large Siemens gas turbines in the United States and South Korea. Sterzing recalls the first test: “We were surprised at how much better the gas turbines could be run this way.” Continuous fine adjustments in the fuel valves optimize how the gas turbines run in terms of emissions and wear by permanently seeking out the best solution in real-time. “To ensure that a gas turbine runs optimally, you always have to search for a balance in which several undesired effects such as combustion dynamics, loss of efficiency and emissions are kept as low as possible. If you improve one variable, you will worsen a different one. Artificial intelligence knows how to find the sweet spot,” Sterzing says.

The gas turbine is just the beginning. Sterzing and Udluft have already used their learning software to control a wind farm. Turbulent flows created by a wind turbine reduce the efficiency of the next wind turbine behind it. This effect is felt throughout the entire wind farm. “You cannot analytically calculate this,” Sterzing emphasizes. That made it an ideal case for learning software.


Transparent Solutions with Interpretable Learning


Deep neural networks (deep learning) has led to rapid advancements in pattern recognition, including in areas such as image and speech recognition. However, the results of their thousands of interwoven, non-linear equations are not generally transparent. “Many new applications, such as steering in autonomous vehicles, or automated decisions in the banking and insurance sector, require comprehensible rules, even if only for legal reasons,” Sterzing explains. In addition to their work on neural networks, therefore, his team is pushing ahead with research on interpretable learning. With the aid of genetic algorithms, these programs can also learn to control highly complex systems and machines. Because the system principally involves equations of freely adjustable complexity, the rules are easier to interpret than in deep neural networks.


Millions of Equations are Analyzed and Recombined in the Cloud


“Engineering expertise has been cultivated and maintained at Siemens for more than 170 years,” Sterzing says. “That is what gives us a very detailed view of the systems and allows us to understand underlying correlations and put them to use for the development of optimal control systems.” This domain know-how is useful for parameterizing the very powerful genetic algorithms. It begins with equations that are randomly combined by input variables. “Initially, this produces a great deal of unusable solutions. But some of these solutions generate minimally better results than the rest. In the next step, the superior solutions are recombined and again produce even better results,” explains Daniel Hein, an Ernst-von-Siemens scholar and doctoral candidate in Sterzing’s research group. As in the process of evolution, the best solutions are created bit by bit. What emerges is a control program with transparent rules that can produce results comparable to that of a neural network. Getting to this point is at least as involved as modeling and training a deep neural network. Millions of equations are analyzed and recombined in the cloud on numerous parallel servers for weeks. But it is certainly faster than evolution in nature, where living things genetically adapt themselves to their environment over millions of years.


Neural Networks and Interpretable Learning Hand in Hand


In Sterzing’s view, deep neural networks and interpretable learning both have their merits, depending on the application. Combining the two methods is also possible. For instance, data models for system identification – to depict how a system such as a wind farm operates – are generated by computing processes in a neural network. But control solutions can also be produced by interpretable learning. Sterzing is confident that this know-how in artificial intelligence will bear fruit in many places along the value chain, whether in service, improving efficiency or in the design stage.


Siemens produces all sorts of complex systems: traffic control systems, industrial automation systems, medical diagnostic systems. The use of artificial intelligence is possible everywhere. “In the near future, a machine that has already learned from data will be more valuable than a brand-new one,” Udluft says. “No one will buy ‘dumb’ devices and systems anymore when smart ones are available at comparable prices,” Sterzing adds. The future has only just begun!


Katrin Nikolaus

Subscribe to our Newsletter

Stay up to date at all times: everything you need to know about electrification, automation, and digitalization.