AI-Assisted Thermographic and Visual Classification of Leading-Edge Erosion of Wind Turbine Blades

by Michael Stamm

This article was originally published in the NDE Outlook column in the?June 2024 issue of?Materials Evaluation. NDE Outlook focuses on possibility thinking for NDT and NDE. Topics may include technology trends, research in progress, or calls to action. To contribute, don't hesitate to get in touch with Associate Technical Editor Ripi Singh at [email protected]

The Importance of NDT on Wind Turbine Blades

The wind industry is a critical element in achieving carbon neutrality. Wind turbines are getting larger and larger, with rotor blades exceeding 100 m in length. As with all machinery, wind turbine blades are regularly inspected to detect and monitor damage. Despite the increasing use of drones and ground-based camera systems, most blade inspections are performed by industrial climbers and the process remains largely manual. Traditional visual inspections help detect damage on the blade surface but cannot capture subsurface information or visualize airflow dynamics. As the wind industry continues to evolve, bridging the gap between traditional visual inspections and more advanced thermographic methods is essential to ensure the efficiency, safety, and sustainability of wind turbines.

One particular type of damage to rotor blades plays a key role in reducing overall efficiency. This is erosion damage to the leading edge of the blades, caused primarily by rain and hail. The significant efficiency loss is caused by changing the aerodynamic properties of the blade before the damage becomes structurally critical. Industry experts estimate that efficiency losses of up to 3% and more can result from erosion damage to the leading edge of rotor blades. Thermographic inspections can visualize and characterize this altered airflow due to leading-edge damage.

Using Thermography and AI Models to Inspect Wind Turbine Blades

The Federal Institute for Materials Research and Testing (BAM) in Berlin has started a research project with partners from industry to economically classify leading-edge damage on rotor blades and to estimate the yield loss due to rain erosion.

Thermographic images of rotating blades taken from the ground are used to visualize the changes in airflow caused by leading-edge damage. The ability of thermography to visualize the effects of leading-edge damage is well known and demonstrated in several scientific papers. It is important to note that the thermographic images do not show the damage itself, but the resulting break in laminar flow.

BAM has been working for years on ground-based thermography of wind turbine blades. For this reason, a specially developed semiautomatic measurement system with a longwave infrared (LWIR) camera was used for the ground-based inspection. A scan of all three blades in operation takes about three minutes and provides 20 to 40 thermographic images, depending on the length of the blade.

In the first phase of the project, 1500 thermographic images of rotor blades with and without visible damage were manually evaluated, annotated, and used to train artificial intelligence (AI) models. This work was carried out by the AI start-up LATODA in cooperation with BAM. One result of the 2023 feasibility study is that a convolutional neural network is well able to detect and classify the flow signatures of leading-edge damage. The classification and detailed calculation of the area is important for correlating the power loss of a wind turbine with this type of damage.

In parallel to this work, LATODA has trained convolutional neural networks to detect and classify leading-edge damage on rotor blades using conventional visual images from drone inspections as well as from ground-based measurements. In both cases, good image resolution is required, which is a particular challenge for ground-based measurements.

Using Neural Networks to Inspect Wind Turbines

The feasibility study has shown that flow effects due to leading-edge damage can be classified by AI-based analysis of thermographic inspections. On this basis, further investigations are currently underway at BAM. The medium-term goal for the end of 2024 is to generate a reference dataset for training neural networks.

To this end, measurements are currently being carried out on 30 wind turbines in Germany. The turbines are from different manufacturers and of different ages, representing a broad spectrum of damage and severity. The dataset contains thermographic inspection data of these turbines, as well as high-resolution visual images. These images are taken at the same time as the thermographic images. For this purpose, the patented visual inspection system of Romotioncam GmbH is used. The advantage of this system is that a rotating camera follows the rotor blades during operation, ensuring minimal blurring of the images. This means that visual and thermographic images are taken simultaneously. This allows for the simultaneous identification of blade damage on the visual images and the corresponding flow signatures on the thermographic images using a single neural network. This comparison promises to provide an understanding of the relationship between leading-edge damage and flow changes.

The open question is whether it is possible to predict stall based on visual inspection. If this is successful, it may mean that thermographic measurements are redundant in terms of flow visualization and will not be needed in the future.

However, a further step is required to calculate the actual performance loss due to leading-edge damage. The performance of individual rotor blades must be calculated from the flow behavior of the air at the blades.

A first attempt to do this is being made with simulations based on work done by Professor Christian Bak from the Technical University of Denmark. His model uses the location and severity of damage on a blade, the main output of the AI model, and applies aerodynamic efficiency-loss factors to each damaged section of the blade and then simulates the overall efficiency loss for the blade.

If successful, the efficiency losses can be correlated with the visual images and valuable operational information can be obtained from relatively simple and inexpensive ground-based visual inspections.

In this project, two physical inspection methods (visual and thermographic) will be implemented with an AI-based evaluation software on different wind turbines and thus different operators. A secure and structured data exchange between the different partners is required to make such a procedure suitable for industrial use in the future.

For this purpose, a data platform according to the European Gaia-X guidelines has been developed and is already in operation. AI concepts such as federated learning can be realized through a trustworthy connection to other data platforms. In this way, AI systems can be improved with data from different operators without having to exchange sensitive data between market competitors.

In addition, the data can easily be merged with other data systems that reflect, for example, operational data and electricity feed-in data. Completely different data systems containing weather data or information on ecosystems can also be merged with the technical operating data, opening up a wide range of analysis options. Which weather conditions lead to the most severe blade erosion, and in which geographical regions is it best to protect the leading edge of the rotor blades?

Author

Michael Stamm: Bundesanstalt für Materialforschung und -prüfung (BAM) (Federal Institute for Materials Research and Testing), Richard-Willst?tter-Str. 11, 12489 Berlin, Germany; [email protected]