Reduced-Order Model Based on Neural Network of Roll Bending
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript presents a neural network-based reduced-order modeling approach for roll bending control, which aligns well with the special issue's focus on intelligent control algorithms. The topic is of great interest and has industrial application merits. However, the technical soundness and quality of presentation require enhancements to fully convey the research contribution and meet the journal's standards. The author is highly recommended to consider the comments below to improve the presentation of this work.
- Introduction:
The form of presentation of the literature review on ROM requires improvement. The author is suggested to demonstrate in the form of statements and analyses in depth instead of simple bullets. Classification of the ROM methods is not applied later. The author should clarify the connection with the following statements on the rolling process. In addition, there are overlaps in the parallelism list, such as “Data-driven modeling, neural networks, machine learning, and deep learning [2–6]”, the latter 3 belong to data-driven modeling, and neural networks are approaches of machine learning and deep learning.
Abbreviations should be defined only once when first showing in the text, e.g., page 2 line 91, “ artificial neural network (NN)” -> (ANN) or artificial NN, since NN has been defined above; page 3 line 107, FEM is not defined.
The introduction should better differentiate this work from existing ROM applications in rolling process control instead of listing and describing the literature.
- Finite Element Model
There should be an overall methodology description in the front of the second section, so that it would be easier for readers to follow the objectives of this work and the approaches to achieve the goals.
It is better to introduce the important quantities of interest in the text instead of in captions, such as in Figure 1.
The first two sentences of the paragraph following Figure 1 seem incomplete. In addition, this section doesn’t describe the 2D and 3D FE models clearly. It looks like only a 2D FE model is presented (depicted in Figure 2), the 3D model is not specified.
Plots in Figure 3 and Figure 4 could use a single legend instead of two legends, and there are missing units in the figures.
Results in Section 2.2 are not clarified based on the 2D or 3D model.
The criteria to decide which simulation model to use are not clear. It is better to validate the model with experimental data and then provide a suggestion to use a 2D or 3D model. After all, the 3D model takes only 2 minutes, which is acceptable for the current computational capability.
Minor edits: page 2 line 78, “rial-time ” -> real-time.
- Neural Network
The criteria to select a neural network are not clear. What are the tolerances of mean square error and accuracy? Is 75% accuracy sufficient for the ROM?
The purpose of Figure 10 is not clear. Please clarify.
The data used in Figure 12 is not clearly stated. The work and analysis of Figure 12 are confusing. Why do we need an NN model to build a â–³Q vs â–³F relationship? The original data should also have â–³Q and â–³F. Do they have the same trend? Make sure whether this relationship is created by the use of a neural network.
Tables should include computational time comparisons under identical hardware conditions to compare the speed of using ROMs with the FE models.
Minor edits: units on page 10, “KN” -> “kN”; line 313, “and a of half of maximum value” is not clear.
- Discussion
The discussion of limitations appears overly brief given the model's industrial application claims.
Minor edits: page 11 line 349, “Strategy of 3D FE calculations. ” is not a complete sentence.
Comments on the Quality of English LanguageThis paper's English is functional, but paying attention to consistency in tense, conciseness, and parallelism in lists, etc., would further enhance its readability and quality. Minor edits are suggested to elevate clarity for journal publication:
- Occasional tense inconsistency of mixing present and past tense in methodology descriptions. E.g., "The model was trained using data from experiments, and it shows high accuracy." -> "The model is trained using data from experiments and shows high accuracy." (Use present tense for general truths); "This paper presents accurate finite element models for a four-high mill, which serve to obtain accurate solutions..." -> "serves".
- Avoid overly complex sentences and redundancy. E.g., "Due to the fact that the deformation of the roll is nonlinear, which is caused by the uneven distribution of stress, it is necessary to consider multiple factors." -> "The roll’s nonlinear deformation, caused by uneven stress distribution, requires consideration of multiple factors."; "In order to achieve the goal of improving accuracy, we..." -> "To improve accuracy, we...".
- Inconsistent use of terms. E.g., "The roll bending process is analyzed... Later, roller bending parameters are optimized." -> "roll bending" vs. "roller bending", should stick to one term (e.g., "roll bending") throughout; "Wider sheets cause less stress and deformation (flattering) on the surface..." -> "flattering" should be "flattening" based on previous descriptions in the paper.
- Missing commas in complex sentences. E.g., "When the load increases the displacement changes nonlinearly." -> "When the load increases, the displacement changes nonlinearly."; "It can be assumed that the shape of the sheet in the output plane of the rolling mill is identical to the shape of the surface of the work roll." -> "It can be assumed that the shape of the sheet in the output plane of the rolling mill, is identical to the shape of the surface of the work roll."
Author Response
Responses in the attached file.
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsIn the paper, a roll bending reduced-order model (ROM) based on neural networks is proposed for real-time control of flatness and convexity in four-high mill hot/cold rolling processes. The effectiveness and superiority of the proposed model are validated by finite element (FE) simulations and industrial data analysis. The reviewer has the following comments:
- The training set contains 3,575 sets of data, but does not specify the data sampling strategy (such as whether it covers the full parameter range or whether there is data bias)..
- The depth of comparison with the existing ROM methods is insufficient. The introduction lists various model order reduction methods (such as POD and RBM), but does not make a detailed comparison of the advantages of the neural network ROM in this paper over other methods in terms of computational efficiency and accuracy. It is suggested to add quantitative comparisons (such as the error curve and calculation time comparison with POD-ROM) to enhance innovation..
- The author did not describe the specific structure of the constructed neural network. It is suggested to add it for reproduction.
- Insufficient consideration of engineering constraints in the bending force model. The flexural force network limits Q to 0-2500 kN, but does not analyze the robustness of the model beyond this range. It is suggested to add boundary condition tests (such as the prediction error when Q=3000 kN) and discuss the setting of safety thresholds in engineering applications.
- It is suggested to discuss the limitations of the proposed method. The author's current research results are merely based on the verification of simulation data. It is necessary to consider how to test the effectiveness of the proposed method in actual engineering applications and whether fine-tuning is needed based on actual data.
Author Response
Responses in the attached file.
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper presents accurate finite element models for a four-high mill, which serve to obtain accurate solutions of the problem of roll bending. The results presented in this paper seem correct, the issue of this study is of practical significance. To improve the paper, the following specify comments are recommended to be considered.
- The figures in the whole manuscript are of high quality, and vector graphics are used for display. The reviewer has a small question, why does Figure 3(a) not give the legend of $2D$ and $3D$?
- The expression in the article needs to be optimized, and it is recommended to use the expression of Figure 12(b).
- The article uses a neural network to train the system model, but lacks an introduction to the neural network architecture used? There are many types of neural network structures. Why did the author use this one? What are the advantages compared to others? Did the author make any innovations?
- Based on the ability of neural networks to approximate unknown functions, they are widely used in the modeling and control of electromechanical systems. It is recommended that the author expand the application of neural networks in the abstract, such as Adaptive neural network-based fixed-time control for trajectory tracking of robotic systems
- The innovation of this article needs to be strengthened, and the existing expression is difficult to publish. The paper is good overall but needs to be strengthened.
Author Response
Responses in the attached file.
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe article is devoted to the development of reduced-order models for control systems of rolling mills. The work has significant practical value. Moreover, the combination of finite element modeling with neural networks as reduced-order models is a promising direction.
However, in my opinion, the author should pay attention to the following points:
Since the main idea of using neural networks for reduced-order modeling is not novel, it is advisable to clearly define the novelty of the author’s approach, particularly in comparison with the methods discussed in the reviewed literature.
It is recommended to perform cross-validation of the neural networks and to add an analysis of their sensitivity to data noise.
The work would benefit from validation of the models using real production data, at least partially.
It is appropriate to provide a reasoned explanation of the chosen neural network architecture and to include an evaluation of their performance compared to baseline models.
Comments on the Quality of English LanguageThe text is generally understandable; however, it contains some grammatical inaccuracies that could be corrected to improve its academic tone.
Author Response
Responses in the attached file.
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have addressed all concerns, and this manuscript should be published.
