Stage-Based Remaining Useful Life Prediction for Bearings Using GNN and Correlation-Driven Feature Extraction
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a novel method for stage-based remaining useful life (RUL) prediction of rolling bearings using a combination of Graph Neural Networks (GNNs) and correlation-driven feature extraction techniques. Experiment results using the IMS bearing dataset demonstrate the model’s accuracy and robustness compared to other approaches. However, there are some places in the paper that can be improved for better clarity and reproducibility:
This paper used a threshold of 0.75 for the Pearson-Spearman correlation and 0.025 for variance in feature selection. It would be more rigorous if the author could explain more about these parameters and their effect on the results.
Although the construction logic for the adjacency matrix is mentioned, the choice of the specific number of neighbors and the weight distribution formula are missing. It is suggested that the author includes a detailed algorithm or pseudocode for the adjacency matrix construction.
To enhance the comprehensiveness of the introduction, I recommend citing a paper addresses challenges in RUL prediction for CFRP structures [1], which share similarities with those in the degradation of bearings, such as handling complex physical properties, limited labeled data, and nonlinear feature dependencies. Adding this reference would provide a broader perspective on the importance of advanced methodologies in RUL prediction, emphasizing the generalizability of the challenges tackled in your research. You could incorporate this reference in the introduction (around lines 73–78), after discussing the limitations of existing methods.
[1] Fatigue Life Prognosis of Composite Structures Using a Transferable Deep Reinforcement Learning-based Approach. Composite Structures.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript presents a study on stage-based Remaining Useful Life prediction for bearings, utilizing a novel approach that integrates correlation-based feature extraction with Graph Neural Networks. The authors propose a methodology that combines the Pearson-Spearman metric, Kernel PCA, Autoencoder, and a GCN-LSTM network to address the challenges of accurately modelling bearing degradation stages and predicting RUL. Validation on the IMS-bearing dataset demonstrates the proposed model's practical applicability and technical robustness. However, this reviewer has identified several areas where the manuscript could be improved:
1. Acronyms such as HI and RMSE must be defined in the text before their first use to ensure clarity for readers.
2.: The research gap is not clearly articulated. Lines 100-104 mention the gap briefly but fail to provide a detailed explanation. This section should be expanded to highlight the novelty and necessity of the proposed approach.
3. The introduction lacks sufficient discussion on LSTM networks, which are central to the methodology. Including this information would provide better context for the readers.
4. The steps in Figure 2 should include directional indicators to improve clarity and understanding of the workflow.
5. It is standard practice in studies to compare the proposed model with other machine learning models to compare performance. While the manuscript describes solid and well-known machine learning models, the novelty is limited. A comparison with alternative models would enhance the paper’s quality.
6. More detailed information on the hyperparameters of the autoencoder latent space and other characteristics of the AI models is required for the possible reproducibility of the project outputs. The manuscript should provide more details about the training and test datasets as well.
7. There are several typos in the paper for example in line 436, which should be corrected.
8, There is an error with one of the references
9. Figure numbering is inconsistent. For example, Figure 4 is referenced as Figure 3 in the text and captioned as Figure 1. This should be corrected.
10. The manuscript does not provide sufficient information on the factors listed in Table 9. A discussion of their differences, advantages, and rationale for selection is necessary as they seem not significantly distinct from each other.
11. However it is true that a separate conclusion is not mandatory, the discussion section in its current form is inadequate. It should be rewritten to provide more data regarding findings, their implications, and potential future work.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for addressing my previous concerns in the revised version of the paper. The updates comprehensively address the points I raised, and I find the current version to be prepared and suitable for publication.