Machine learning and Big Data have increased competitiveness in many industries by being able to predict certain behaviours. The Eurostars WINDELIN project began with focusing on the detection of ice on a turbine blade, with a view to reducing wind farm maintenance costs.
“Current sensors tend to apply pure physics to detect icing. What we wanted to do was gather tons and tons of data to find patterns that show us when something is beginning to behave abnormally,” explains Ivan Tyagov, WINDELIN project manager and head of European R&D projects for Nexedi, France. “This enables the operator to do something about the situation.”
In many northern countries, long fences (of roughly 390m) must be built around turbines to ensure that falling ice does not cause harm. This is expensive and an inefficient use of land. Early detection of icing could help operators know when to stop operations before there is a risk of danger. “If an abnormality is detected early, it is also much cheaper to repair a blade before a whole turbine needs to be dismantled,” adds Klaus Wölfel, head of Nexedi’s German subsidiary. “If the turbine is vibrating unnaturally over a long period of time, this can also damage the motor.”
While the WINDELIN project focused on ice detection, the innovation developed can be applied to other structural health monitoring tasks. “We can scale this project horizontally towards sectors like construction and oil and gas,” says Tyagov. “Wherever there is need to detect structural health, this technology has huge potential.”
Applying machine learning
The WINDELIN project brought together three partners with complementary knowledge and expertise. “Cooperation was crucial,” says Tyagov. “In modern engineering, having a strict segregation between various components makes the job virtually impossible. Tight integration was needed from all partners: from the field hardware engineers to the software developers, and finally the data science team. All this would have been impossible without the collaborative effort.”
The first key milestone was the manufacture of a prototype sensor: a small cubed box that can be placed inside a wind turbine to collect vibration data. From a technical point of view, this was challenging due to the harsh conditions that wind turbines typically face.
The next tasks involved finding out how to store the massive amounts of data collected and developing machine learning algorithms to detect anomalies. “Getting access to data is extremely hard, as this is a heavily regulated business,” says Wölfel. “The value of data is high.” The WINDELIN project was able to access data from a wind farm in Germany and run field tests.
“After some initial evaluation work, we focused on developing the data platform,” says Tyagov. “During this phase, we discovered that even state of the art algorithms could not cope with the vast amount of incoming data from wind turbines. We had to find ways of reducing the amount of data required, and eventually incorporated ideas used in the Large Hadron Collider (LHC) to pre-process some data directly at the sensor.”
Exploring industrial applications
The final result is ready for commercialisation. The project developed a fully functional platform that integrates sensors, data, wind turbine management and machine learning in an all-in-one subscription service for wind farm operators. “This plug-and-use solution can be tailored to specific needs and scenarios,” says Wölfel. “Subscription revenues are shared among project partners through a revenue sharing agreement.”
A customised WINDELIN solution is currently being used by a large wind turbine service operator in Germany. “At the moment, this solution collects and analyses data from more than 450 wind turbines in close to real-time,” says Tyagov.
This focus on distributed data gathering and processing (where machine learning is built into the sensor itself) reflects the direction that the industry is moving in. “More calculations are being done, whether in the factory or in the turbine, by using cloud computing,” says Tyagov.
The machine learning-based innovation is also context-dependent. A sensor installed in one turbine cannot work in another, because the environment is different. “This is interesting from a business point of view, as we would like to export this technology to China,” Tyagov says. “While physical algorithms can be copied, machine learning that is unique to each sensor cannot.”
Both Wölfel and Tyagov expect to duplicate their success in other fields where Big Data analysis and smart sensor management can bring cost benefits. “We initially only targeted wind turbine operators and farms,” says Wölfel. “But any industry where the automation of detection and control of physical processes is needed has potential.” This includes the oil industry, the automotive and aviation industries as well as industrial automated manufacturing in general. Nexedi is also looking into construction, where sound and vibration data could be collected to ensure health and safety standards and structural integrity.
Partners: Nexedi SA (France), Micromega Dynamics SA (Belgium), Mariadb Corporation Ab (Finland)
Project ID: 9483 WINDELIN (Eurostars)
Project duration: January 2015 to June 2018
