Machine Learning Finds Defects in Wind Turbine Blades

Machine Learning Finds Defects in Wind Turbine Blades

AI-based system finds signal of damage within noisy radar data.
Wind turbines can be difficult to maintain, especially offshore. Getting out to sea to inspect the turbines is a dangerous and cost-intensive job, and engineers have been working to make the parts, specifically the turbine blades, more durable to reduce maintenance. Recently, a collaborative team of researchers from the University of Glasgow and École Polytechnique Fédérale de Lausanne (EPFL) found a novel way to test the blades for defects that may increase their longevity.

“If we can drive down the costs and improve the performance of the wind turbine blade, which needs to withstand a really dynamic and hostile environment, then we can make a significant positive impact,” said David Flynn, professor in cyber physical systems at Glasgow’s James Watt School of Engineering. “We looked at the fundamental issue of how can we detect and ensure the quality of these blades, not just at the point of manufacture, but in their operational usage as well.”

Turbine blades are complex structures made up of several composite layers, and the collected signals are noisy, which makes it difficult to find irregularities. Olga Fink, assistant professor of civil engineering and head of the Intelligent Maintenance and Operations Systems Laboratory (IMOS) within EPFL’s School of Architecture, Civil and Environmental Engineering (ENAC), has been working on algorithms that can decipher this noise. Her team’s previous projects have involved detecting and predicting faults anomalies in machinery, suppressing background noise in different types of condition monitoring signals, including audio, strain, and vibration recordings, and also distinguishing bird species through song recordings.

The new method employs a robotic arm to inspect the blades at distances as large as 15 cm. Before now, the further you were from the blade, the more difficult it was to detect the anomalies.
With this new method, Fink explained that they introduced new complex-valued activation functions within their machine learning algorithms that leverage the characteristics of data collected by frequency modulated continuous-wave radar. By removing the healthy data features, the team was able to focus on the irregularities in the signals that would indicate different anomalies within the blades. They were also able to capture robust data collected at different distances.

“Utilizing a complex-value representation of the signals has enabled us to more effectively separate the information they contain and to adapt the AI model accordingly by integrating the new activation function,” Fink said. “And because we focused not on the signals themselves, but on the differences of those signals from the healthy conditions, we introduced a compact representation of different noise levels.”

Flynn and his team integrated the new method into an existing radar system that employs a robotic arm to inspect the blades at distances of 5, 10, and 15 centimeters. Before now, the further you were from the blade, the more difficult it was to detect the anomalies.

Learn from ASME: Introduction to Computational Fluid Dynamics

“We can now measure from different distances without physically touching the structure,” Flynn said. “And we can scan it and give it that kind of ultrasound-equivalent analysis and localize any defects. This method really is about bringing a lot more clarity to those problems.”

Some of the issues the new method can detect include uncured resin and delamination, two common problems that are below the surface and therefore completely invisible to the human eye. Flynn said it has been common to get anywhere between three to six false alarms when analyzing the structures. Even when the blades are properly inspected and maintained, they still don’t have a long lifespan.

“We're going to learn more about how these structures dynamically perform in near real time, which will be a treasure trove of information,” Flynn said.

The team is eager to get real-world data—not just during the manufacturing process but also from installed wind turbines during operation—using the new method and the next step will be working with service companies to get proof of concept. They want to learn how their technology can provide ease of use to those doing the inspections, especially because the design is so portable.

Fink is eager to see how their methodology can process signals from different distances under harsh and noisy conditions. Her team hopes to continue their efforts in exploring both the spatial dimensions of the signals and the interdependence of the signals, or how the signals are influencing each other.

One of the current research directions of her team is prescriptive operation. “We want to know how we can adjust the operational parameters to prolong the useful lifetime of the systems and components, given the operational requirements,” she said. “This is a very exciting and very promising direction or us.”

More like this: Floating Pyramids Could Make for Cheaper Wind Power

Flynn hopes collecting this data and shoring up their methods will give the industry the ability to more accurately model and create future turbine blades.

“These blades are notoriously hard to recycle,” he said. “They end up in landfills, which contradicts the idea that these are sustainable energy technologies. Hopefully, we can now do more to create a different end-of-life story for this technology.”

Cassandra Kelly is a technology writer in Columbus, Ohio.

You are now leaving