the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Active Trailing Edge Flap System fault detection via Machine Learning
Andrea Gamberini
Imad Abdallah
Abstract. Active trailing edge flap systems (AFlap) have shown promising results in reducing wind turbine (WT) loads. Once the WT design includes the AFlap, a condition monitoring system will be needed to ensure the flaps provide the expected load reductions. This paper presents two approaches based on machine learning to diagnose the health state of an AFlap system. Both approaches rely only on the sensors commonly available on commercial WTs, avoiding the need and the cost of additional measurement systems. The first approach (MFS) uses manual feature engineering in combination with a random forest classifier. The second approach (AFS) relies on random convolutional kernels to create the feature vectors. The study shows that the MFS method is reliable in classifying all the investigated combinations of AFlap health states in the case of asymmetrical flap faults not only when the WT operates in normal power production but also before startup. Instead, the AFS method can identify some of the AFlap health states for both asymmetrical and symmetrical faults when the WT is in normal power production.
- Preprint
(1521 KB) - Metadata XML
- BibTeX
- EndNote
Andrea Gamberini and Imad Abdallah
Status: final response (author comments only)
-
RC1: 'Comment on wes-2023-24', Anonymous Referee #1, 23 Jun 2023
Dear editor,
I have completed a thorough review of the manuscript titled "Active Trailing Edge Flap System fault detection via Machine Learning." After careful evaluation, I would like to provide my feedback on this manuscript.
The manuscript analyzed an important topic in the wind turbine field/industry, which is leveraging machine learning methods to identify the health state of active trailing edge flaps (AFlaps) on wind turbines. The study explores two methods, namely the manual feature selection (MFS) and automatic feature selection (AFS) approaches. The ML models were trained and tested using simulated data.
Overall, the topic and analysis of the manuscript is of a great interest of the wind energy science/industry. The procedure and results of the research is well organized and presented. The performance of the two methods and final conclusions are also well delivered.
However, I have several concerns that need to be addressed before considering the manuscript for publication:
- Lack of clarity regarding the problem addressed: The abstract and introduction do not clearly articulate the specific real-world problem or practical application that the research work aims to solve. I can see that the main goal of the two models is to predict the health state of AFlaps with the accessible signals without installing additional sensors. However, this main goal is not stated in the abstract. Also, the current stage (e.g., the performance of similar research on this topic) of related research effort is not presented in the introduction. I recommend providing a clear statement of the problem, presenting the previous reported research, and highlighting the potential impact or benefit of solving it.
- Incomplete analysis of AFS method performance: As reported in the manuscript, the AFS method works very well for the AF_Off and AF_Off_Fault cases, which shows the potential of the AFS method in accurately distinguishing the flap on-off states. In this case, people might wonder why the AFS cannot do well in the opposite cases (i.e., AF_On and AF_On_Fault) which basically differ from the AF_Off and AF_Off_Fault cases only by the Flap actuator control signal. I recommend having further analysis/investigations to answer the following questions: i) What are the potential reasons why the model works worse on the AF_On and AF_On_Fault cases? ii) Will putting the Flap actuator control signal as a feature into the AFS method increase the performance?
- Potential improvement: As reported in the manuscript, the MFS method showed reliable classification for asymmetrical flap faults but struggled with symmetrical flap faults, and the AFS method had limited success in both cases but outperform in two special symmetrical cases. I recommend having an investigation on integrating the two model into a hybrid model since combining different models/techniques can often lead to improved performance and enhanced capabilities, leveraging the strengths of each individual method.
I believe addressing these concerns and incorporating the suggested improvements will strengthen the manuscript and enhance its contribution to the field. Overall, the research work has done a great contribution to advance the understanding of fault detection in active trailing edge flap systems and improve the reliability of wind turbine operations.
Thank you for considering my feedback. I am available for any further clarifications or discussions.
Sincerely
Citation: https://doi.org/10.5194/wes-2023-24-RC1 -
AC1: 'Reply on RC1', Andrea Gamberini, 10 Aug 2023
Dear Anonymous Referee #1,
Thank you for your comments.
- Lack of clarity regarding the problem addressed:Â We agree that the specific real-world problem and benefits need to be clearly stated. We have now modified the text in the abstract (lines 1 to 5 and 12 to 14), introduction (lines 24 to 34), and conclusion (lines 573 to 575) to add more clarity to the subject matter. Regarding previous research on the topic, we could not find any relevant papers. Therefore, in the introduction (lines 34 to 58), we described three approaches commonly used for detecting and monitoring faults in wind turbine components.
- Incomplete analysis of AFS method performance:Â The actuator control signal is already included in the input signal used for the generation of the features (line 176)
- Several parameters can affect the performance of the AFS. Due to time and resource constraints, we want to show that the AFS method has some potential and, in some configurations, performs better than MFS. However, it needs to be further investigated and developed to be usable. We added this consideration in the AFS discussion (lines 562 to 568) and the conclusions (lines 607 to 609).
- Potential improvement:Â Thank you for the comment.[IA2] Ensemble modeling is a technique used by the authors of this paper and can confirm the reviewer's comment that it very well might be an ideal approach to fusing the strength of multiple methods simultaneously. We will note this improvement in our future research. Due to time and resource constraints, it cannot be evaluated in this current version of the paper. The idea has been added as possible future development (lines 616 and 617).
Citation: https://doi.org/10.5194/wes-2023-24-AC1
-
RC2: 'Comment on wes-2023-24', Davide Astolfi, 16 Jul 2023
The paper deals with a very interesting topic and is appropriate for the scientific objectives of the journal.
The objective of the work is stated extremely clearly. The quality of the work is high, the methods are applied rigourously. The presentation is excellent.
Therefore, my recommendation is to accept this paper upon the following minor revisions:
- I am interested in a few further details about the type of employed data. What is the time resolution of the data?
- Do the authors think that the same kind of objective can be accomplished for on site wind turbine with standard SCADA data averaged on ten minutes? Could you please elaborate a little more on this?
Citation: https://doi.org/10.5194/wes-2023-24-RC2 -
AC2: 'Reply on RC2', Andrea Gamberini, 10 Aug 2023
Dear Davide Astolfi,
Thank you for taking the time to review our work. Here are our responses to your questions:
- The data are obtained from aeroelastic simulations of 10 minutes with a time resolution of 0.01 s. ( line 137). We developed this methodology considering that it should be applicable to commercial wind turbines. The Manual Feature Selection approach with a reduced set of features relies only on 10 minutes statistical properties of commonly available wind turbine signals. Therefore, we believe this approach can be directly implemented in an actual prototype. The model must be trained with simulations based on the target wind turbine aeroelastic model and eventually tuned with transfer learning techniques using the wind turbine SCADA data. The MLS method with full features requires calculating additional features generally not included in the standard SCADA data. A cost-benefit evaluation should be performed to decide which features are relevant to be captured in addition to the SCADA data. The Automatic Features Selection methodology does not use SCADA data. It requires instead the special postprocessing of the 10 minutes high-frequency sampled time series. We included this consideration in the chapters of the discussion(lines 546 to 551 and 569 to 571) and conclusion (lines 599 to 604 and 611 to 612 ).
Citation: https://doi.org/10.5194/wes-2023-24-AC2
Andrea Gamberini and Imad Abdallah
Andrea Gamberini and Imad Abdallah
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
531 | 114 | 16 | 661 | 7 | 6 |
- HTML: 531
- PDF: 114
- XML: 16
- Total: 661
- BibTeX: 7
- EndNote: 6
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1