Structural Damage Features Extracted by Convolutional Neural Networks from Mode Shapes
Round 1
Reviewer 1 Report
The range of applications of neural networks is rapidly expanding and the works analyzing the details of their functioning are especially valuable. The present paper reports a study of the efficiency of convolutional neural network in damage detection in model steel truss based on simulated data as well as gives some insight into the neural network functioning.
The model system is a steel truss consisting of 355 rods, each of which can be damaged (what is quantified by reduction of its elastic modulus in the range 2…95%). The Authors calculate the first order mode shapes for all the cases of the damage in the system. Either the mode shape or the mode curvature (second derivative of the shape) is taken as the input parameters for the convolutional neural network. The performance of the neural network is compared for both these cases of the input data.
The idea of the paper is clear, the presentation and analysis of interesting results is generally convincing and the topic is of importance for technical sciences.
I recommend the manuscript for publication in Applied Sciences, provided that the Authors consider the following points listed below (related to the presentation of the results):
Line 118: I guess “while” at the beginning of the sentence is not necessary.
Line 120: I guess the second derivative Φ’’ should also have an index k, being centered at k
Line 121: What is the value of Δl?
Line 150: The caption is partially covered by the figure itself.
Line 169: Is n the number of classes? (also appears in line 340) – this would be explained.
Lines 177-178 and lines 196-197: The same sentence with “….the training process for 1, 2, 3 and 4 convolutional layers” would sound more clear.
Line 218, figure 7: There is some problem (I guess a technical one) with the figure: it presents only a single mode, whereas the legend and the sentence in line 216 refers to more modes.
Figure 8 a,b,c,d: Can the position of the damaged rod (the range on the horizontal axis) be marked? This would facilitate correlation the mode curvature difference with the position of the damaged element (as stated in line 224).
Figures 11, 12, 13 and 14: I guess it would be better from the point of view of comparison of the curves to plot them all in a single plot, so to supplement Fig. 15 with one more curve and remove separate plots shown in Figs. 11-14. The common plot would need an appropriate legend with keys like A,B,C,D explained in the caption, like “A: The difference between the average of training samples and the sample of the 3202th scenario” and so on). This way the tiny differences would be better observable.
Also separate Figs. 16 and 17 are not necessary because of Fig. 18 showing both curves together (once more, just a change in the legend/caption would be sufficient, with removing separate Figs. 16 and 17).
Author Response
Revisions and Explanations
Manuscript Number: applsci-822945
Title: Structural damage features extracted by convolutional neural networks from mode shapes
Authors: Kefeng Zhong, Shuai Teng, Gongfa Chen*, Gen Liu, Fangsen Cui,
The authors are grateful for the reviewers' efforts in improving this paper. In response to the reviewers’ comments, the specific modifications are as follows:
Responses to Reviewer 1
- “The range of applications of neural networks is rapidly expanding and the works analyzing the details of their functioning are especially valuable. The present paper reports a study of the efficiency of convolutional neural network in damage detection in model steel truss based on simulated data as well as gives some insight into the neural network functioning.
The model system is a steel truss consisting of 355 rods, each of which can be damaged (what is quantified by reduction of its elastic modulus in the range 2…95%). The Authors calculate the first order mode shapes for all the cases of the damage in the system. Either the mode shape or the mode curvature (second derivative of the shape) is taken as the input parameters for the convolutional neural network. The performance of the neural network is compared for both these cases of the input data.
The idea of the paper is clear, the presentation and analysis of interesting results is generally convincing and the topic is of importance for technical sciences.
I recommend the manuscript for publication in Applied Sciences, provided that the Authors consider the following points listed below (related to the presentation of the results):”
Response: Thank you very much for your encouraging comments.
- “Line 118: I guess “while” at the beginning of the sentence is not necessary.”
Revision: Revised as suggested.
- “Line 120: I guess the second derivative Φ’’ should also have an index k, being centered at k.”
Revision: Revised as suggested.
- “Line 121: What is the value of Δl?”
Revision: Line 121 is changed to “Δl = 0.05 m, it is the distance between two adjacent points.”
- “Line 150: The caption is partially covered by the figure itself.”
Revision: Revised as suggested.
- “Line 169: Is n the number of classes? (also appears in line 340) – this would be explained.”
Revision: Yes, we added the sentence “n is the number of classes” in line169, and the n in Line 340 is the power factor of the power function, we added a sentence “n = power factor”
- “Lines 177-178 and lines 196-197: The same sentence with “.the training process for 1, 2, 3 and 4 convolutional layers” would sound more clear.”
Revision: Revised as suggested.
- “Line 218, figure 7: There is some problem (I guess a technical one) with the figure: it presents only a single mode, whereas the legend and the sentence in line 216 refers to more modes.”
Revision: Some first order mode shape samples are shown in Figure 7. There are 14 curves of different damage scenarios in Figure 7 and these curves have different colors; but they are so similar that they look like only one curve.
- “Figure 8 a,b,c,d: Can the position of the damaged rod (the range on the horizontal axis) be marked? This would facilitate correlation the mode curvature difference with the position of the damaged element (as stated in line 224).”
Revision: Revised as suggested.
- “Figures 11, 12, 13 and 14: I guess it would be better from the point of view of comparison of the curves to plot them all in a single plot, so to supplement Fig. 15 with one more curve and remove separate plots shown in Figs. 11-14. The common plot would need an appropriate legend with keys like A, B, C, D explained in the caption, like “A: The difference between the average of training samples and the sample of the 3202th scenario” and so on). This way the tiny differences would be better observable.”
Revision: Revised as suggested. We deleted the Figure 13 and 14. But the Figure 11 has a different dimension (Figure 11 contains 143 pieces of data, while Figure 12, 13 and 14 contain 195 pieces of data), so we drew it separately and Figure 12 is retained for comparison. We also added some legends.
- “Also separate Figs. 16 and 17 are not necessary because of Fig. 18 showing both curves together (once more, just a change in the legend/caption would be sufficient, with removing separate Figs. 16 and 17).”
Revision: Revised as suggested.
Author Response File: 
Author Response.doc
Reviewer 2 Report
This paper explains about the structural damage features extracted by convolutional neural networks from mode shape. The title of this manuscript appears to be interesting to peers, but the content of this work looks somehow without innovative contributions to the related (well established) methodologies. The author should make an effort to extend his idea to real problems. This reviewer was impressed with the amount of data that went into this research. That alone gives one hope that a paper worthy of the archival standing of a major academic journal such as this is possible. The paper cannot be accepted in the present form as it needs further improvements.
-Abstract: The text must be carefully revised. Some sentences contain mistakes (in the abstract: very general statements) whereas some sentences must be reworded as the English is “meaningless”. I strongly recommend that the authors retain the services of a professional editor. There are many reputable companies that offer these services.
- Introduction is poorly written. Proper references need to be used rather than using others. Language can be improved. The sentences are half constructed or incomplete in a way that the readers are expected to fend for themselves in order to understand their meaning.
-I suggest the authors to review lot more papers in the following journals for more important contents, namely, CNN, SHM techniques, mode shape extraction techniques. Also, try to add a section which explains about the experimental part separately and tabulate the results. It will be very helpful for the peer researchers.
- The innovation contribution of this article is not clearly stated. The research contributions should be highlighted in the revised manuscript.
-Authors must be enriching the references with some latest developments in the field. Some of the recent references can be added. It is evident the authors have not paid attention to previous research papers and concerns.
The list could go on, but the bottom line is that the authors need to rewrite the paper, or even reconsider the research content, before it could be considered for publication in this journal. There are lot of minor errors that need to be addressed. It is difficult to mention all here. Few errors are,
(1) Try to explain in detail about the significance / key points in utilizing different finite element approaches. What happen to the non-linear problems?
(2) Almost every figure need to be updated. It is of very poor quality. When Matlab plots are used, try to use bigger text font. It is very difficult to understand.
(3) Change Figure 2 with proper representation. In Figures 7-19, mention the units. Also, Figures 9 and 10 are not very clear to understand.
(4) Rewrite discussions and conclusions section. Well known statements can be removed and only discuss about the research content.
What is more, the paper is not well written, and difficult to understand at points. Due to the fact that this work fails to provide convincing innovation and lacks in depth viewpoints, I would recommend ‘Major revision’ for this paper.
Author Response
Revisions and Explanations
Manuscript Number: applsci-822945
Title: Structural damage features extracted by convolutional neural networks from mode shapes
Authors: Kefeng Zhong, Shuai Teng, Gongfa Chen*, Gen Liu, Fangsen Cui,
The authors are grateful for the reviewers' efforts in improving this paper. In response to the reviewers’ comments, the specific modifications are as follows:
Responses to Reviewer 2
- “This paper explains about the structural damage features extracted by convolutional neural networks from mode shape. The title of this manuscript appears to be interesting to peers, but the content of this work looks somehow without innovative contributions to the related (well established) methodologies. The author should make an effort to extend his idea to real problems. This reviewer was impressed with the amount of data that went into this research. That alone gives one hope that a paper worthy of the archival standing of a major academic journal such as this is possible. The paper cannot be accepted in the present form as it needs further improvements.
-Abstract: The text must be carefully revised. Some sentences contain mistakes (in the abstract: very general statements) whereas some sentences must be reworded as the English is “meaningless”. I strongly recommend that the authors retain the services of a professional editor. There are many reputable companies that offer these services.
- Introduction is poorly written. Proper references need to be used rather than using others. Language can be improved. The sentences are half constructed or incomplete in a way that the readers are expected to fend for themselves in order to understand their meaning.
-I suggest the authors to review lot more papers in the following journals for more important contents, namely, CNN, SHM techniques, mode shape extraction techniques. Also, try to add a section which explains about the experimental part separately and tabulate the results. It will be very helpful for the peer researchers.
- The innovation contribution of this article is not clearly stated. The research contributions should be highlighted in the revised manuscript.
-Authors must be enriching the references with some latest developments in the field. Some of the recent references can be added. It is evident the authors have not paid attention to previous research papers and concerns.
The list could go on, but the bottom line is that the authors need to rewrite the paper, or even reconsider the research content, before it could be considered for publication in this journal. There are lot of minor errors that need to be addressed. It is difficult to mention all here. Few errors are:”
What is more, the paper is not well written, and difficult to understand at points. Due to the fact that this work fails to provide convincing innovation and lacks in depth viewpoints, I would recommend ‘Major revision’ for this paper.
Response: Thank you for your criticisms and advices, in order to state the innovation contribution of this article clearer, we have made several changes to the abstract and introduction section, language has been improved and more references have been updated.
- “Try to explain in detail about the significance / key points in utilizing different finite element approaches. What happen to the non-linear problems?”
Revision: This paper adopts an elastic model without plastic deformation, damage is defined as changes to the material elastic modulus, which adversely affect the current or future performance of these systems.
- “Almost every figure need to be updated. It is of very poor quality. When Matlab plots are used, try to use bigger text font. It is very difficult to understand.”
Revision: We have updated some figures (Figure 2, 8-18) to make it clearer and easy to be understood.
- “Change Figure 2 with proper representation. In Figures 7-19, mention the units. Also, Figures 9 and 10 are not very clear to understand.”
Revision: Revised as suggested. But the data in Figures 7-19 are dimensionless, they are generated in the data processing of the CNN.
- “Rewrite discussions and conclusions section. Well known statements can be removed and only discuss about the research content.”
Revision: We deleted line 306 to 311 “In essence, the CNN is a mathematical fitting ···· ···· it will have a better ability to present a complex function” and made some revision in the conclusion 3.
But most other parts of the discussions and conclusions of the original manuscript are based on the research content of this paper; we think they are necessary and significant, so we kept the original version.
Author Response File: 
Author Response.pdf
Reviewer 3 Report
This paper consider an availability of CNN for locating damage rods in a complex steel truss. The theme is cutting edge and very interesting. I judged that this paper is acceptable after minor revision. My detailed comments and questions are listed below:
- What is the definition of “damage”? It should be clarified in the paper.
- In Fig. 2. The figure of truss (green rods) is shifted.
- In Fig. 2. What is the blue shape in the upper right? If it is not needed, it should be delete.
- In Fig. 9. There are 16 curves in each condition. What is the difference between these curves?
Author Response
Revisions and Explanations
Manuscript Number: applsci-822945
Title: Structural damage features extracted by convolutional neural networks from mode shapes
Authors: Kefeng Zhong, Shuai Teng, Gongfa Chen*, Gen Liu, Fangsen Cui,
The authors are grateful for the reviewers' efforts in improving this paper. In response to the reviewers’ comments, the specific modifications are as follows:
Responses to Reviewer 3
- “This paper considers an availability of CNN for locating damage rods in a complex steel truss. The theme is cutting edge and very interesting. I judged that this paper is acceptable after minor revision. My detailed comments and questions are listed below:”
Response: Thank you very much for your encouraging comments.
- “What is the definition of “damage”? It should be clarified in the paper.”
Revision: In this paper (Line 104): “The structural damage is defined as changes to the elastic modulus of the rod concerned. It is assumed that the damage degree of a rod has a linear relation with the reduction of its elastic modulus.”
- “In Fig. 2. The figure of truss (green rods) is shifted.”
Revision: Revised as suggested.
- “In Fig. 2. What is the blue shape in the upper right? If it is not needed, it should be deleted.”
Revision: Revised as suggested.
- “In Fig. 9. There are 16 curves in each condition. What is the difference between these curves?”
Revision: They are the output data of 16 channels of each layer of the CNN. They are generated in the data processing of the CNN. Each curve represents a feature extracted by the CNN. Obviously, they're different, at least in shape, actually, they are also different in value range, it depends on the internal weights of the CNN, but we normalized the output data in this paper.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors have addressed almost every comment except for improving the figures.
Still, the quality of the figures is very poor and difficult to read. Try to use the proper software to improve the figure illustrations.
Still, there are typo errors in the article, which need to be carefully revised.
Once, these comments are addressed, the article can be published. So I am requesting the editor to give, "Minor revision" in this regard.
Author Response
Revisions and Explanations
Manuscript Number: applsci-822945
Title: Structural damage features extracted by convolutional neural networks from mode shapes
Authors: Kefeng Zhong, Shuai Teng, Gongfa Chen*, Gen Liu, Fangsen Cui,
Responses to Reviewer 2
Comment:
The authors have addressed almost every comment except for improving the figures.
Still, the quality of the figures is very poor and difficult to read. Try to use the proper software to improve the figure illustrations.
Still, there are typo errors in the article, which need to be carefully revised.
Once, these comments are addressed, the article can be published. So I am requesting the editor to give, "Minor revision" in this regard.
Response: We are grateful for your efforts in improving this paper. Sorry for the bad experience caused by these poor quality figures, in response to your comments, we've improved almost all the figures and proofread the article carefully. Figures 1, 3, 7, 9, and 15 have been redrawn and other figures have been greatly optimized.
Author Response File: Author Response.doc
