Remaining Useful Life Prediction Method Based on Dual-Path Interaction Network with Multiscale Feature Fusion and Dynamic Weight Adaptation
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper presents Remaining Useful Life (RUL) prediction by using deep learning methods including multiple convolution kernels of different scales through a 1D Convolutional Neural Network. The paper is interesting but contains significant shortcomings.
4oAbstract-
There is no information about the C-MAPSS dataset. The paper's target is related to the data set. The reader should know what the paper's solved. Which application area? Predictive maintenance in which equipment?
The paper means "existing methods mainly focus on single-scale, single-dimensional feature extraction, which limits the accuracy of RUL prediction". However, there is no information about problem or data type like time series. They assumes that the reader know everything. This is a significant weakness.
The abstract should include some results also.
2. Materials and Methods
Define all variables in the equations. For example what are W and b in (2)?
Define all words in figures. For example, Wwat are FC1-FC4?
Include the recent references also by introducing over them your contribution.
The recent reference are not sufficient.
3. Experiment
Figure 6 and the words on it are not explained in the text.
What is DPI?
the number of DPI layers, and preliminarily verify the effectiveness of the model to explore 403
How did you select parameters of N and W. give some references for it.
N = [2, 4, 6, 8, 10] and W = [10, 20, 30, 40, 50, 60].
Give some benchmark results and its the recent reference to compare with your results in anew table.
Discuss new results.
Remove multiple abbreviations. Some examples are:
21-
demands of efficient operations[2]. By leveraging Remaining Useful Life (RUL) prediction,
52
resulting in more accurate Remaining Useful Life (RUL) predictions. Currently, the more
In this paper, we proposed a novel framework for remaining useful life (RUL) predic- 533
----
widely used models are Recurrent Neural Networks (RNNs) and Convolutional Neural 53
Networks (CNNs) , RNNs excel at capturing temporal features, while CNNs are effective 54
-
both accuracy and interpretability[18]. Chen et al. utilized convolutional neural networks 89
(CNNs) to extract temporal features and combined them with a Bayesian long short-term 90
----
proposed a parallel hybrid neural network combining 1D convolutional neural networks 110
(1-DCNN) and bidirectional gated recurrent units (BiGRU) to simultaneously extract both 111
BTSAM[17]: A hybrid neural network that combines 1D Convolutional Neural Net- 430
works (1-DCNN) with Bidirectional Gated Recurrent Units (BiGRU) in parallel. 431
Figure and Table styles are well given.
Conclusion
Give your limitations. Evaluate the computational load of the proposed model.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsDear Authors,
The manuscript presents a new approach for Remaining Useful Life (RUL) prediction within the field of Prognostics and Health Management (PHM). The Dual-Path Interaction Network is proposed to facilitate the interaction between temporal and feature dimensions. The network integrates different modules to enhance feature extraction and fusion across multiple scales and dimensions. The proposed method is validated using the C-MAPSS dataset, demonstrating superior performance compared to existing SOTA methods in terms of RMSE, MAPE, and Score metrics. I believe the paper is interesting, and overall I found the analysis quite robust and well-conducted. That said, I have some comments, listed below.
1) Methodologies. The description of DPIN lacks sufficient detail on how the temporal and feature paths interact beyond element-wise summation. Can you please add more information on the interaction mechanisms? Also, regarding the MTF-CFM module, how was the specific size of the convolution kernel chosen? Some discussion regarding how the concatenation and splitting operations contribute to feature richness would also be a nice addition.
2) Experiment section. Can you provide more information and justification for the selection of the specific 14 sensors? Also, what was the rationale behind the chosen hyperparameter ranges (in particular, for window sizes and DPI layers)?
3) The results in Table 3 are presented with average improvements, but there are no statistical tests (such as t-tests) to assess the significance of performance gains. Since DL methods have stochastic components in their training process (for example, coming from the weights initialization and the optimization process) showing the average result of multiple runs (with in brackets the standard deviation) might be helpful to strengthen the validity of performance gains.
4) Discussion section. While you mention the need for computational efficiency improvements, there is no information in the paper regarding the proposed model’s computational requirements (such as training time, memory usage, etc.) compared to the baselines. Since you acknowledge this, it is more a curiosity of mine, but a table with such information might be helpful for the readers, especially those who would like to apply similar techniques in real-world use cases.
Comments on the Quality of English LanguageOverall, the English language quality is good.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsMy comments are addressed.
Reviewer 2 Report
Comments and Suggestions for AuthorsDear Authors,
Thank you very much for your revisions. Overall, I am satisfied with them. Thanks again for the interesting research, and I wish you good luck.
Comments on the Quality of English LanguageEnglish quality is overall good.
