Armed with learning algorithms, virtually any highly complex system can be designed to minimize its own maintenance requirements and improve its output. Examples include advanced medical equipment, power distribution systems, gas turbines and entire wind parks.
CT’s Amit Chakraborty has developed learning software that can predict power requirements. Machine learning could play a role in expanding a smart grid in Allgäu, Germany.
Cross-section of a gas turbine. Using neural networks, learning systems can predict the optimal operating criteria for turbines and their associated emission levels.
In the future, Personal Energy Agents will handle power trading between consumers and power companies via a specialized box equipped with learning software.
Siemens researchers are currently testing the system.
Learning from Sounds - and Saving a Lot of Energy
open
- Text Size
- Share
- Print this page
Hollywood likes to play with the idea of intelligent robots. Just think of the autonomous and uncontrolled machines in the blockbuster film “Transformers.” But reality is different. Most moviegoers probably don’t know that researchers have already made great strides in giving machines the ability to learn and act on their own — always for the benefit of humanity, of course.
This is the sort of work that’s being done at Siemens Corporate Technology (CT) in Princeton, New Jersey. There, a team working with Knowledge Decision Systems Program Manager Amit Chakraborty is developing a new type of software for power companies. The software can learn the habits of electricity customers by analyzing millions of data records. Eventually, the system will be able to independently make forecasts of power demand. In the “smart grid” of the future, the main objective will be to reconcile power consumption with fluctuating sources such as solar and wind power plants, for example. “Sustainable energy systems will then manage consumer load current so as to adapt it to the fluctuating generation of current by renewable energy sources,” says Chakraborty. “That’s why we have to develop methods that allow power companies to do precise planning.”
Before the end of 2011, the new software will be tested in a pilot project using real world power consumption data. The objective is, first of all, to study consumers’ energy use profiles. To that end, data will be collected from millions of customers using intelligent power meters. The data will provide information about the quantities of power used and the periods in which it was used. Siemens researchers will combine what they learn from the pilot project with meteorological data and information about special events, such as baseball playoffs. They will use this trove of raw data to develop training data for their software. Its algorithms will then be able to create accurate short-term load forecasts.
Load forecasts are not a new invention. Everyone is familiar with the peak load that occurs during holidays when millions of turkeys disappear into ovens. But these very rough patterns fall short of the requirements for a sustainable energy system. In the U.S., power companies have been relying on market principles for years when managing loads. If a lot of electricity is available, the cost is lowered. Conversely, consumers can enter into contractual commitments to use less power when supplies are tight or pay a higher price for it if they don’t. But “demand response” systems of this kind don’t always work perfectly. If consumers don’t behave as expected, power companies quickly have to produce or purchase additional electricity; this is often inefficient and leads to higher greenhouse gas emissions. “To prevent this, we have to be able to predict how consumers will behave under the conditions that exist at any particular time,” says Chakraborty.
Machine learning could also help to reduce the cost of expanding the electrical grid. For instance, Dr. Michael Metzger, who is researching the automation of power grids for an advanced “smart grid” project at Siemens in Munich, has developed — along with other CT experts — a learning algorithm that calculates the structure of the grid using measurements made by sensors. “Often, no information is available about the number or location of copper lines laid decades ago to supply end users with power,” he says. To get this sort of basic information about hidden parts of the electrical grid, sensors can be placed within the network of cables. These provide data regarding the flow of current and voltage at their locations. Armed with this information, it is possible to determine the structure of the grid. “This information allows a grid operator to know how much voltage its network has and where that voltage is located,” says Metzger. Siemens is currently testing the estimation algorithm in part of the electrical grid of the Allgäuer Überlandwerke power company in Kempten, in southern Germany.
Recognizing the Signs of Failure. In the service field, the changes that machine learning will bring could be revolutionary. Instead of waiting for failures in expensive equipment such as medical diagnostic systems, Siemens researchers are taking a giant step forward. “We’ve developed a program that can reliably predict when a magnetic resonance imaging (MRI) machine or a nuclear medicine system will fail,” says Dr. Fabian Mörchen, who develops learning systems in the Knowledge & Decision Systems area at Siemens’ research site in Princeton. The approach starts with the fact that there are telltale signs in many machines that can indicate when failure is imminent. “The trick is to identify those signals and make them visible,” says Mörchen. Such signals can include changes in electrical currents, voltages, noises, vibrations, pressures, and temperatures.
Deviations from normal operation are measured by sensors in the machines themselves. Based on information regarding what is normal for a machine, researchers and their learning systems use data mining to filter out anomalous patterns. Once a series of patterns has been correlated with a malfunction, the team working with Mörchen can develop algorithms that train a computer program to identify those patterns when processing data that hasn’t been seen before. For example, when the cryogenic helium in an MRI scanner begins to escape there are very slight changes in temperature and pressure. Thanks to early warning algorithms, technicians at Siemens Healthcare can zero in on the problem and repair the cooling system before the machine fails. Today, Siemens service teams use this software not only to monitor more than 3,500 MRI scanners but also to perform preventive maintenance. This strategy has resulted in a reduction in maintenance and repair costs of $5.8 million over a period of three years.
One of the forerunners of these research projects was a program led by CT Researcher Ciprian Raileanu in Princeton that was designed to monitor bridges. At the time, the U.S. Department of Transportation was looking for a way to optimize the maintenance and repair of the roughly 650,000 bridges in the U.S. Raileanu’s team developed a solution in collaboration with Rutgers University and its Center for Advanced Infrastructure and Transportation, which is located near Princeton.
“The system uses data from sensors at the bridges, inspection reports, weather data, historical data from construction diagrams, accident frequencies from police reports, and photographs to independently determine a bridge’s condition,” says Raileanu. “We examined this very heterogeneous data for patterns,” he adds. Based on these patterns, algorithms learn what consequences might result from the convergence of certain factors. For instance, if the bridge was built in 1976 in a region with heavy precipitation and has iron girders, it is very probable that cracks have formed in the piers after 30 years. The U.S. Department of Transportation has been using the bridge-monitoring program since 2008.
The program also served as a model for an entirely new system that railway companies in Great Britain and Russia are using to monitor their train fleets. The data for this learning software comes from sensors in various subsystems in trains, such as those that monitor brakes and doors, as well as from train schedules and fault reports. Known as the Rail Remote Service Desktop (RRSD), the system combines all this data and calculates where each train is at any given time and whether maintenance work is needed. At present, RRSD is monitoring 175 trains — and Siemens supplies not only the software but also the automation components.
Mastering Complexity. Gas turbines are another major application area for learning systems — in this case, those based on neural networks (see From biological Systems to Machines Learning is the key and Pictures of the Future, Spring 2011, Turning Many into One). These systems create forecasts regarding emissions and optimal turbine operation in a matter of seconds . Turbines are governed by innumerable complicated interrelationships that researchers can often only assess via statistical methods, since many values can only be roughly estimated. Traditional mathematical formulas requiring exact figures are thus not very useful in this research. But to maximize a turbine’s lifespan and performance while minimizing its emissions the effects of thousands of settings have to be precisely assessed and forecast.
Volkmar Sterzing and his CT team at Siemens’ Intelligent Systems & Control Global Technology Field in Munich have therefore developed a new method that makes this possible. Using so-called recurrent neural networks, the researchers can depict a turbine’s entire processes and thus make accurate forecasts regarding its output. “Previously, you could only get a snapshot of these process,” explains Sterzing. “In effect, our new method allows us to take pictures of what’s happening before and after that snapshot.” According to Sterzing, this method enables researchers to know not only what happened in the past but also how processes will continue in the future. This dynamic representation makes it possible to identify and make the most of changes that are positive while reducing the impact of those that may be negative and altering maintenance plans accordingly.
CT researchers have applied what they have learned from gas turbines to a related field — the optimization of wind turbines and entire wind parks. As an ardent regatta sailor, Sterzing knows that in order to steer his boat in the best possible way he must keep an eye on the waves, wind speed, and competing sailboats during every minute of a race. Otherwise, it would not be possible to forecast future developments and plan an appropriate course. This approach inspired him to create a software system for wind turbines based on sensors that measure about ten factors, including wind speed, turbulence levels, temperature, and air pressure. Algorithms correlate this data with a wind park’s output so that the software can learn from thousands of interrelationships and apply its knowledge to novel situations.
As the system learns different situations, it gets better and better at independently forecasting which settings — such as the rotor blades’ angle of incidence or the generator speed — are required for a wind turbine to generate the greatest output from the available wind. This method has been show to increase a wind turbine’s output by up to half a percentage point. That may not sound like much, but in a large wind park it has a major impact. Tests at the Lillgrund wind farm in Sweden during the last six months have shown that, thanks to the ability to learn independently from its own actions — so-called autonomous learning — the park has been able to increase its output by the equivalent of an additional turbine.