A Fault Early Warning Method Based on Auto-Associative Kernel Regression and Auxiliary Classifier Generative Adversarial Network (AAKR-ACGAN) of Gas Turbine Compressor Blades
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript introduces a fault warning technique for gas turbine compressor blades based on AAKR-ACGAN. I recommend accepting the manuscript upon resolving the following issues.
(1) Some key quantitative results can be presented in Abstract.
(2) The authors should include and discuss more newly published fault diagnosis methods, such as generalized fault diagnosis framework based on phase entropy, and high-accuracy intelligent fault diagnosis method for aero-engine bearings.
(3) Some figures are blurry and should be replaced with high-resolution versions, such as Figures 1, 2, 9, 10, 16.
(4) The corresponding spectrum results can be provided in Figure 12, which help confirm the effectiveness of wavelet denoising.
(5) It is recommended to optimize Figure 16, as the relevant information is difficult to read. Additionally, conclusions derived from Figure 19 should be included to supplement the analysis.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsGeneral Comments:
This study proposes a fault warning technique for gas turbine compressor blades based on AAKR-ACGAN. First, a digital twin model of the gas turbine is constructed using long-term operation data and simulation data from the mechanism model. Then, an Auto Associative Kernel Regression (AAKR) model is used for fault warning, monitoring multiple parameters to provide effective early warnings of potential faults. Additionally, an Auxiliary Classifier Generative Adversarial Network (ACGAN) is employed to fully extract hidden data features of fault points, balance the dataset, and accurately simulate the process of fault occurrence and development. The proposed approach is utilized for the early detection of faults in the compressor blades of a high-capacity gas turbine and its precision and applicability have been confirmed.
Specific Comments:
1. Page 1, line 34-37: the statement about the compressor blades failure is very critical to this article and should be properly cited “overtime, these blades...and property”:
o Khan et al. (2021). Centrifugal Compressor Stall Control by the Application of Engineered Surface Roughness on Diffuser Shroud Using Numerical Simulations. Materials, 14(8), 2033. https://doi.org/10.3390/ma14082033
o Advanced materials and technologies for compressor blades of small turbofan engines
o Blade roughness effects on compressor and engine performance—a CFD and thermodynamic study
2. Page 3, line 92-94: There are a lot of input parameters discussed here, what inputs are used for regression analysis and how those inputs have been chosen?
3. Page 5, line 152: it is mentioned that for compressor mechanism modelling, mechanistic chamber modelling is used. In general, Mechanistic chamber model requires a lot of internal details including the geometry of the compressor, frictional and heat losses, etc. I don’t see any of those models considered here. Please explain how the mechanistic chamber model was used.
4. Page 5, line 173: what is fault injection? Please explain it in text for a better understanding of the readers.
5. Figure 4, 5 and 6: Please add updated high pixels image.
6. Page 11, section 2.4.1: why mean squared error was used instead of mean absolute percentage error? I think MAPE has better interpretability and comparability than MSE.
7. What were the hyperparameters for the model explained above? Please explain as I don’t see it in the article. It would be better if you could incorporate it somewhere in the methodology section.
8. Figure 12: I don’t see any visual difference between “before denoising” and “after denoising”? Please explain
9. Page 16, line 487-498: “Ultimately, 6 linearly...fault warning”, what are those 6 variables?
10. Table 5: what was the training/testing data set percentage? I mean how much data was used for training and how much was used for testing?
11. Figure 22: I don’t see the mean line for MMSPE, it is hard to conclude what your data represents. Please include that mean line for better understanding.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
