the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 Jan 2023
Regression-based Main Bearing Load Estimation
Abstract. The premature failure of wind turbines due to unknown loads leads to a reduction in competitiveness compared to other energy sources. Especially the failure of main bearings results in high costs and downtimes, as for an exchange of this component the rotor needs to be demounted. Load monitoring systems can make a significant contribution to understand and prevent such failures. However, most load monitoring systems do not take into account the main bearing loads in particular as there is no commercially applicable measuring system for this purpose. This work shows how main bearing loads can be estimated using virtual sensors. For this purpose, several regression models are trained with test bench data considering strain and displacement signals. It is investigated with which combination of signal type and regression model the highest accuracy is achieved. The results show that for either using strain or displacement signals an appropriate accuracy can be achieved. In particular, it is shown that a linear regression with interactions already achieves good accuracy and that further increases in regression model complexity do not add significant value.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(935 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2022-75', Anonymous Referee #1, 01 Feb 2023
The submitted article with number wes-2022-75 entitled ‘’ Regression-based Main Bearing Load Estimation’’ presents an experimental study on the uncertainty of measuring wind turbine main shaft loads with strain and displacement sensors. The topic is of scientific relevance and the applied methodology follows best practices and is free of major errors. Significant improvements could be made in the presentation of the methodology and of the results, which does not satisfy journal quality standards.
1. Introduction
1.1. The novelty of your work could be better emphasized. Have similar investigations on displacement-based load measurements with full-scale test benches been conducted by other researchers, or is this study unique in that regard?
1.2. It is stated that the development of displacement-based load estimation systems is still in the concept stage, however, recent works conducted by NREL have demonstrated the applicability of this approach in the field. (Guo, Yi, Bankestrom, Olle, Bergua, Roger, Keller, Jonathan, and Dunn, Mark. Investigation of main bearing operating conditions in a three-Point mount wind turbine drivetrain. United States: N. p., 2021. Web. doi:10.1007/s10010-021-00477-8.)
2. Methodology
2.1. Please specify whether the DUT is an onshore or offshore wind turbine.
2.2. Explain why axial loads could not be measured with strain gauges. Is this a limitation of the testbench, the commercially available sensors, budgetary…? Furthermore, elaborate why the large rotor of the low wind speed turbine corresponds to a low signal to noise ratio in thrust induced strains (line 116).
2.3. It is stated that the load time series is obtained by multibody simulation (line 135). Do you mean aero-elastic simulation instead? Please briefly describe the simulation models and software used here.
2.4. Gravity forces are not considered in the calculation of the bearing reaction forces (Eq. 1-5) and in the analytical model (Eq. 6-9). Are these negligible?
2.5. Please elaborate on the motivation and the benefit of using LRI and LRS. It is clear that higher polynomial terms are needed to capture the progressive bearing stiffness curves, however, it is not reasoned why the coupling terms are necessary. It would be beneficial to provide an example of such variable interdependencies.
2.6. Provide model equations for the machine learning models GPR, RFR and FNN or add reference (e.g. The Elements of Statistical Learning: Data Mining, Inference and Prediction, T. T. Hastie, Robert; Friedman, Jerome, Publisher: Springer 2009. DOI: 10.1007/b94608)
2.7. The performance of the machine learning models largely depends on the selected hyperparameters (number of nodes and layer, activation function,…) and the training function. Please specify your approach.
3. Results
3.1. Contrary to the regression models, the analytical model does not have an intercept term (alpha_0). It should therefore be checked if this model is producing biased results (if the mean values of measurements and applied loads agree).
3.2 The claim ‘’ First, since very good accuracies are already achieved with almost all methods, the costs of the measurement system can be reduced (e.g., usage of less sensors) without significantly losing accuracy’’ is unsubstantiated. The presented results only suggest that high accuracy is achieved using all sensors and do not allow any conclusions on the performance of a subset of sensors. To verify your claims, please add the achievable RMSE and R^2 of your proposed sensor setup (one axial displacement + two bending strain gauges) to tables 2 and 3.
3.3 Since one of your proposed applications is fatigue damage monitoring, it would be informative to add a table with the error in the bearing fatigue damage.
3.4. Please elaborate on the different sources of uncertainty. Is measurement noise an issue? Why is the measurement of radial loads more accurate than axial loads?
4. Technical comments: The manuscript is at times difficult to follow due to grammatical errors and informal language. Please refer to the attached pdf for my recommended language edits.
- RC2: 'Comment on wes-2022-75', Edward Hart, 13 Feb 2023
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2022-75', Anonymous Referee #1, 01 Feb 2023
The submitted article with number wes-2022-75 entitled ‘’ Regression-based Main Bearing Load Estimation’’ presents an experimental study on the uncertainty of measuring wind turbine main shaft loads with strain and displacement sensors. The topic is of scientific relevance and the applied methodology follows best practices and is free of major errors. Significant improvements could be made in the presentation of the methodology and of the results, which does not satisfy journal quality standards.
1. Introduction
1.1. The novelty of your work could be better emphasized. Have similar investigations on displacement-based load measurements with full-scale test benches been conducted by other researchers, or is this study unique in that regard?
1.2. It is stated that the development of displacement-based load estimation systems is still in the concept stage, however, recent works conducted by NREL have demonstrated the applicability of this approach in the field. (Guo, Yi, Bankestrom, Olle, Bergua, Roger, Keller, Jonathan, and Dunn, Mark. Investigation of main bearing operating conditions in a three-Point mount wind turbine drivetrain. United States: N. p., 2021. Web. doi:10.1007/s10010-021-00477-8.)
2. Methodology
2.1. Please specify whether the DUT is an onshore or offshore wind turbine.
2.2. Explain why axial loads could not be measured with strain gauges. Is this a limitation of the testbench, the commercially available sensors, budgetary…? Furthermore, elaborate why the large rotor of the low wind speed turbine corresponds to a low signal to noise ratio in thrust induced strains (line 116).
2.3. It is stated that the load time series is obtained by multibody simulation (line 135). Do you mean aero-elastic simulation instead? Please briefly describe the simulation models and software used here.
2.4. Gravity forces are not considered in the calculation of the bearing reaction forces (Eq. 1-5) and in the analytical model (Eq. 6-9). Are these negligible?
2.5. Please elaborate on the motivation and the benefit of using LRI and LRS. It is clear that higher polynomial terms are needed to capture the progressive bearing stiffness curves, however, it is not reasoned why the coupling terms are necessary. It would be beneficial to provide an example of such variable interdependencies.
2.6. Provide model equations for the machine learning models GPR, RFR and FNN or add reference (e.g. The Elements of Statistical Learning: Data Mining, Inference and Prediction, T. T. Hastie, Robert; Friedman, Jerome, Publisher: Springer 2009. DOI: 10.1007/b94608)
2.7. The performance of the machine learning models largely depends on the selected hyperparameters (number of nodes and layer, activation function,…) and the training function. Please specify your approach.
3. Results
3.1. Contrary to the regression models, the analytical model does not have an intercept term (alpha_0). It should therefore be checked if this model is producing biased results (if the mean values of measurements and applied loads agree).
3.2 The claim ‘’ First, since very good accuracies are already achieved with almost all methods, the costs of the measurement system can be reduced (e.g., usage of less sensors) without significantly losing accuracy’’ is unsubstantiated. The presented results only suggest that high accuracy is achieved using all sensors and do not allow any conclusions on the performance of a subset of sensors. To verify your claims, please add the achievable RMSE and R^2 of your proposed sensor setup (one axial displacement + two bending strain gauges) to tables 2 and 3.
3.3 Since one of your proposed applications is fatigue damage monitoring, it would be informative to add a table with the error in the bearing fatigue damage.
3.4. Please elaborate on the different sources of uncertainty. Is measurement noise an issue? Why is the measurement of radial loads more accurate than axial loads?
4. Technical comments: The manuscript is at times difficult to follow due to grammatical errors and informal language. Please refer to the attached pdf for my recommended language edits.
- RC2: 'Comment on wes-2022-75', Edward Hart, 13 Feb 2023
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 355 | 139 | 35 | 529 | 28 | 28 |
- HTML: 355
- PDF: 139
- XML: 35
- Total: 529
- BibTeX: 28
- EndNote: 28
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1