A Convolutional Neural Network-Based Method for Distinguishing the Flow Patterns of Gas-Liquid Two-Phase Flow in the Annulus

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

 

Comments

1.      The abstract would benefit from being shorter, focusing more on the central findings and implications than on long descriptions of the methodology.

2.      Summary of main findings in terms of percent accuracy for each flow type with an emphasis on how significant it is for real-time monitoring.

3.      Although well-arranged, the introduction is densely written due to the more than a few methods of flow analysis applicable to the case without explicitly dividing them into their different divisions and association with this paper. Therefore, the proper use of subheadings for methods shall improve.

4.      The concluding part of the introduction for the present work should finally convey a stronger statement that explicitly states its contribution to research vis-à-vis any study on the set as its key objective. For this, it ensures a smoother transition toward methods

5.      While describing the CNN architecture, some decisions made in critical parameters like kernel size, pooling, and number of layers could be much better described. For example, one could discuss why 3x3 kernels or a particular activation function was preferred over others.

6.      The steps involved in data preprocessing may also be better described. It would help if the author mentioned whether any noise reduction technique was applied or whether any specific criteria were followed for choosing the image augmentation.

7.      Though the performance of CNN is reported, it doesn't research further why performance is different at places, for example, between slug and annular flows. In that regard one discussion could give some substance to why the accuracy in slug flow is considerably poor compared with that of the annular flow.

8.      It also indicates brief recommendations about how issues related to the recognition of churn flow may be dealt with; however, ideas deserve more elaborated forms, though, to fit in well within a separate "Future Work" section

9.      Some of the graphics, like the model flow chart and the gas-liquid phase distribution, as in Fig. 5, 6, could be enhanced with legends or labeling to make for better readership by those readers without some familiarity with the visual recognition of flow analysis techniques.

10. A tabular summation of details related to CNN training data in the form of sample count per type of flow would add ease of handling data make data behind the process more transparent and allow easy comparison across flow types in performance.

11. This would further boost the discussion comparing the performance of the CNN-based model with other techniques related to flow recognition discussed in the introduction. This will enable the model to strengthen the claim for real-world applications.

12. Applicability of the model to real-time industrial environments in terms of computational cost or hardware requirements could be taken care of as potential limitations of the model.

13. Manuscript requires close editing for correct grammatical errors and thus makes it better readable. Some of the sentences used in methods and results should be shortened to improve the flow of the text in that section. For example, a term like "Sequential model" or "Evaluate function" can be made explicit in the sentence for a non-expert to clearly understand it without knowledge of certain programming or CNN terminologies.

 

14. A 35% similarity report is relatively high, especially for an original research manuscript. Recommend that the authors carefully revise to reduce the similarity score, particularly in non-technical sections.

Comments on the Quality of English Language

Manuscript requires close editing for correct grammatical errors and thus makes it better readable. Some of the sentences used in methods and results should be shortened to improve the flow of the text in that section. For example, a term like "Sequential model" or "Evaluate function" can be made explicit in the sentence for a non-expert to clearly understand it without knowledge of certain programming or CNN terminologies.

Author Response

Comments 1: The abstract would benefit from being shorter, focusing more on the central findings and implications than on long descriptions of the methodology.

Response 1: Abstracts have been reorganized and summarized. See ‘Abstracts’ in the manuscript for details.

 

Comments 2: Summary of main findings in terms of percent accuracy for each flow type with an emphasis on how significant it is for real-time monitoring.

Response 2: Thank you so much for such insightful comments. The text marked in yellow and red in the introduction to the manuscript has been supplemented. This study will provide theoretical support for the online identification methods of gas-liquid two-phase flow patterns in annuli, thereby enriching the theoretical framework of multiphase flow pattern recognition.

 

Comments 3: Although well-arranged, the introduction is densely written due to the more than a few methods of flow analysis applicable to the case without explicitly dividing them into their different divisions and association with this paper. Therefore, the proper use of subheadings for methods shall improve.

Response 3: The introduction has been organized and supplemented with the addition of relevant images to deepen the understanding of the research context. See the labelling in the introduction for details.

 

Comments 4: The concluding part of the introduction for the present work should finally convey a stronger statement that explicitly states its contribution to research vis-à-vis any study on the set as its key objective. For this, it ensures a smoother transition toward methods

Response 4: The last paragraph of the introduction is rewritten to not only summarize the results of the previous work, but also state the shortcomings of it, leading to the research methodology that will be adopted in this paper as well as the research objectives that are intended to be achieved, so as to make it serve as a bridge between the previous and the next. See page 2 of the manuscript for details.

 

Comments 5: While describing the CNN architecture, some decisions made in critical parameters like kernel size, pooling, and number of layers could be much better described. For example, one could discuss why 3x3 kernels or a particular activation function was preferred over others.

Response 5:  To address the complexity and uniqueness of slug flow, this study employs 3×3 convolutional kernels, which offer high parameter efficiency, strong feature extraction capabilities, and ease of stacking. Compared to larger kernels, the 3×3 design possesses fewer parameters, thereby reducing model complexity and mitigating the risk of overfitting, effectively extracting edge and texture features from images. The architecture comprises three convolutional layers with 32, 64, and 64 kernels respectively, each tasked with extracting key features from the input feature images through convolutional operations. The model utilizes the ReLU activation function, which excels in converting linear inputs into nonlinear outputs, enhancing the model's classification performance. Additionally, ReLU's simplicity contributes to high computational efficiency. Added to section 2.3.1 of the manuscript.

 

Comments 6: The steps involved in data preprocessing may also be better described. It would help if the author mentioned whether any noise reduction technique was applied or whether any specific criteria were followed for choosing the image augmentation.

Response 6: The specific image preprocessing process is:

1. Data cleaning:

Initial screening of the data was first performed to remove incomplete, anomalous and duplicate records.

1. Data standardization/normalization:

In order to eliminate the effect of different magnitudes, the numerical type features were standardized or normalized.

1. Noise Reduction Processing:

In image data preprocessing, median filtering noise reduction technique was ap-plied to reduce the image noise and improve the image quality.

1. Feature engineering:

For image data feature extraction and selection is done, including features such as color, texture, shape, etc.

1. Image Enhancement:

While selecting the image enhancement techniques, it is ensured that the enhanced image needs to be realistic, no additional noise is introduced, and the enhancement effect significantly improves the model performance. Specific methods: rotation, scaling, flipping, contrast adjustment, etc., to ensure that the model has a better generalization ability to the transformed image. Specific steps for image pre-processing have been added to the manuscript, as detailed on page 7.

 

Comments 7: Though the performance of CNN is reported, it doesn't research further why performance is different at places, for example, between slug and annular flows. In that regard one discussion could give some substance to why the accuracy in slug flow is considerably poor compared with that of the annular flow.

Response 7: This study reveals that slug flow is primarily caused by unstable gas-liquid flow rates, leading to blurred gas-liquid interfaces in flow pattern images. Even after initial image processing, it remains challenging to eliminate the instability of the gas-liquid interface caused by flow fluctuations. This blurriness is a major impediment to the recognition accuracy of Convolutional Neural Network (CNN) models. In contrast, annular flow exhibits a clear gas-liquid interface, where the liquid and gas phases can flow stably within the pipe and maintain a fixed flow pattern, resulting in a high recognition accuracy for annular flow. Detailed at the end of section 3.4 of the manuscript.

 

Comments 8: It also indicates brief recommendations about how issues related to the recognition of churn flow may be dealt with; however, ideas deserve more elaborated forms, though, to fit in well within a separate "Future Work" section

Response 8: Thank you for your comment. We have made appropriate modifications to the conclusion section. Detailed at the end of section 4 of the manuscript.

 

Comments 9: Some of the graphics, like the model flow chart and the gas-liquid phase distribution, as in Fig. 5, 6, could be enhanced with legends or labeling to make for better readership by those readers without some familiarity with the visual recognition of flow analysis techniques.

Response 9: Gas-liquid interface labelling has been added to the corresponding figures. See page 12 for details.

 

Comments 10: A tabular summation of details related to CNN training data in the form of sample count per type of flow would add ease of handling data make data behind the process more transparent and allow easy comparison across flow types in performance.

Response 10: Tables 1 and 2 have been reorganized as detailed on pages 11-12.

 

Comments 11: This would further boost the discussion comparing the performance of the CNN-based model with other techniques related to flow recognition discussed in the introduction. This will enable the model to strengthen the claim for real-world applications.

Response 11: By comparing our findings with those of Zhao et al., we note that despite both studies employing neural network models to predict gas-liquid two-phase flow patterns, there is a significant difference in prediction accuracy. Zhao's research indicates that the flow pattern recognition accuracy of the traditional BP neural network is 87.5%, while the SSA-optimized BP neural network improves the recognition accuracy to 91.66%. In our study, the trained CNN model achieved prediction accuracies of 0.9265 for slug flow and 0.9873 for annular flow, respectively. This demonstrates that the model employed in our study has a clear advantage in terms of prediction accuracy. See page 13 for details.

 

Comments 12: Applicability of the model to real-time industrial environments in terms of computational cost or hardware requirements could be taken care of as potential limitations of the model.

Response 12: The model does have relatively high requirements for real-time industrial environments in terms of computational cost or hardware requirements. As a reminder for future researchers, a hint is given in the last sentence of the conclusion of the manuscript. See the conclusion for more details

 

Comments 13: Manuscript requires close editing for correct grammatical errors and thus makes it better readable. Some of the sentences used in methods and results should be shortened to improve the flow of the text in that section. For example, a term like "Sequential model" or "Evaluate function" can be made explicit in the sentence for a non-expert to clearly understand it without knowledge of certain programming or CNN terminologies.

Response 13: Thank the reviewer for the good comment, and the authors have polished our manuscript carefully and corrected the grammatical, styling, and typos found in our manuscript. The long sentences in the manuscript have been shortened.

Comments 14: A 35% similarity report is relatively high, especially for an original research manuscript. Recommend that the authors carefully revise to reduce the similarity score, particularly in non-technical sections.

Response 14: The authors have rewritten the parts of the paper with high repetition rates, and the manuscript has been checked for plagiarism using Turitin. The overall similarity rate is less than 20%, and all similar sources are less than 1%.

 

Comments on the Quality of English Language

Manuscript requires close editing for correct grammatical errors and thus makes it better readable. Some of the sentences used in methods and results should be shortened to improve the flow of the text in that section. For example, a term like "Sequential model" or "Evaluate function" can be made explicit in the sentence for a non-expert to clearly understand it without knowledge of certain programming or CNN terminologies.

Response: Thank you for your valuable and thoughtful comments, the authors have carefully checked and improved the English writing in the revised manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

1. In Section 1, please add some introductive photos, such as the actual industrial applications and relevant harms of undesired flows.

 

2. Please add primary assumptions/hypotheses of this work at obvious places.

 

***3. Results and discussion are only given in Section 4.3, which is very short and requires to be significantly enriched for journal publication. 

 

4. Please add an uncertainty analysis of the research.

Comments on the Quality of English Language

The writing could be proofread for improvements.

Author Response

Comments 1: In Section 1, please add some introductive photos, such as the actual industrial applications and relevant harms of undesired flows.

Response 1: Thank you for the reminder to add relevant images in the ‘Introduction’ to improve the comprehensibility of the article. In addition, the Introduction has been further summarized and a title has been added. See the yellow labelling in the Introduction（31-36） and Figure 1 for details.

Comments 2: Please add primary assumptions/hypotheses of this work at obvious places.

Response 2: This item was overlooked during the writing of the manuscript and relevant assumptions about the model have been added in section 2.3.2 (lines 256-258).

Comments 3: Results and discussion are only given in Section 4.3, which is very short and requires to be significantly enriched for journal publication. 

Response 3: After an in-depth analysis of the structure of the article, the overall structure of the manuscript has been reorganized. In order to further improve the logical soundness of the article, the ‘Results and analyses’ section has been reorganized. See Section 3 (pp. 8-13) for details.

Comments 4: Please add an uncertainty analysis of the research.

Response 4: ‘Uncertainty analysis’ has been added to the manuscript, see page 10-11 for details.

Comments on the Quality of English Language

The writing could be proofread for improvements.

Response: Thank you for your valuable and thoughtful comments, the authors have carefully checked and improved the English writing in the revised manuscript.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed all major comments in detail, with adjustments and clarifications made as recommended, especially in the abstract, introduction, and CNN methodology. The remaining suggestions—such as adding a few sentences about  "Limitations and Future Work" in conclusion, improving figure descriptions, and including a brief glossary would further polish the manuscript without requiring significant content revisions.

Author Response

Comment：The authors have addressed all major comments in detail, with adjustments and clarifications made as recommended, especially in the abstract, introduction, and CNN methodology. The remaining suggestions—such as adding a few sentences about  "Limitations and Future Work" in conclusion, improving figure descriptions, and including a brief glossary would further polish the manuscript without requiring significant content revisions.

Response: Thank you for your thoughtful comments, the authors have carefully revised the conclusion section，figure description has been improved and a brief glossary has been added at the end of the revised manuscript.  See the red labelling in the manuscript for details.

Author Response File: Author Response.pdf