the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Comparative anomaly detection for floating offshore wind turbines using in-situ data
Abstract. This study introduces a comparison between a kurtosis-analysis and a deep learning-based approaches for detecting anomalies of a floating offshore wind turbine. The study uses in-situ measurements from a 2.3 MW floating offshore wind turbine named Zefyros and deployed approximately 11 kilometers off the coast of Norway. The first method employs the statistical metric of kurtosis to detect anomalies within a signal by identifying variations in the signal distribution. The second method employs a deep-learning procedure based on an auto-encoder approach, which transforms inputs into a reduced-dimensional latent space and then uses the encoded information to produce outputs identical to the inputs. One month of SCADA and high frequency measurements obtained thanks to S-Morpho sensors were used in the study. Due to limitations in the accessible SCADA information, the anomaly scenario was simplified to detecting whether the turbine rotor was rotating or not. Both tested methodologies can accurately detect anomalies, with the ground truth based on rotor rotations per minute (rpm) measurements. The auto-encoder method shows promising results, delivering more accurate outcomes than the kurtosis analysis on this in-situ measurement dataset. This study is a first step toward a more general use of auto-encoders for wind turbine anomaly detection. The latent space build by the auto-encoder can be leveraged to detect other types of anomalies, with a few labeled data.
- Preprint
(10228 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on wes-2024-189', Anonymous Referee #1, 03 Apr 2025
reply
Good manuscript overall, could use some improvements and expand on some sections, please see comments below:
- Line 135, not sure this sentence makes sense
- On the methodology section, it would be useful to have a flowchart showing the two methods followed, which sensors they used etc
- Line 250: can you further elaborate on the filtering process? were there any curtailments to consider?
- I understand that sensor location has been explained in another paper, but the authors could consider including a summary on their placement and sensor characteristics would be useful for the reader.
- Figure 9: Further clarification on which sensor each signal comes from would and what processing was followed would be be useful
- Line 275, how was the specific period of magnetic signal selected?
- Line 276: "The kurtosis values were then compared to SCADA values", can you elaborate in more detail how this comparison was done?
- Line 280: Can you explain or cite a source that explains the kurtosis categories?
- Line 285: can you further explain the connection between kurtosis thresholds and operating states?
- Line 299: which weather numerical prediction models?
- Line 308: Why were there two different activation functions used in different layers?
- Figure 13: Can you comment on which data is beyond the threshold in the training phase? Were the non operating conditions used in the training?
- Line 333: Typo, "a statistical method"
- A bit more discussion on the results would be good. Do the authors recommend using ANN? Do they recommend bigger time windows for kurtosis estimation? did they do a sensitivity analysis on this?
- Also what type of anomalies can they detect with this methodology?
- Additionally it would be nice to visualise the inference on the test/unseen data and how the model could detect anomalies in real time if implemented
- How common is it to not have SCADA over the sensors used in the paper? Are they widely available in commercial wind turbines? is this method more useful for field trials?
Citation: https://doi.org/10.5194/wes-2024-189-RC1
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
100 | 31 | 5 | 136 | 10 | 7 |
- HTML: 100
- PDF: 31
- XML: 5
- Total: 136
- BibTeX: 10
- EndNote: 7
Viewed (geographical distribution)
Country | # | Views | % |
---|---|---|---|
United States of America | 1 | 45 | 33 |
France | 2 | 18 | 13 |
United Kingdom | 3 | 12 | 9 |
China | 4 | 11 | 8 |
Germany | 5 | 7 | 5 |
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1
- 45