An Improved Target Network Model for Rail Surface Defect Detection

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposed an improved target network model for rail surface defect detection. Here are my comments:

 

1. In the Introduction, the authors said “Traditional manual inspections and contact measurement technologies persist in fault detection within rails. However, these methods not only require a significant number of engineers and are time-consuming and labor-intensive”. However, many studies have proven that contact measurement technology can detect rail defects through onboard detection methods, such as those as follows. Please refer to these references to modify the manuscript:

Rail Crack Detection Using Optimal Local Mean Decomposition and Cepstral Information Coefficient Based on Electromagnetic Acoustic Emission Technology

CUFuse: Camera and Ultrasound Data Fusion for Rail Defect Detection

2. In the second paragraph of page 2, the authors listed some NDT methods, such as ultrasonic detection, eddy current, laser detection, and acoustic emission. The acoustic emission technique is missing. Please summarize the advantages and disadvantages of acoustic emission based on "EMAE-Based Rail Structural Health Monitoring Using Double-Layer Signal Processing and Spectrum Information Entropy".

3. This paper uses deep learning algorithms to detect rail surface defects. This method requires annotation of training images using labeling in the data preparation stage. However, it is difficult to obtain labeled data in the actual detection site. Does this deep learning-based detection method have limitations? How can it be solved?

4. Please re-summarize the contribution of this paper, because YOLOv7 and k-means are existing algorithms, and it seems to be a simple fusion of different algorithms.

5. In recent years, some scholars have also used the YOLO model to detect rail surface defects. Please compare their methods to verify the effectiveness of this method, instead of just comparing different network models in the submitted manuscript.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Please include a description of the abbreviations before using them for the first time e.g. line 65, line 134

2. Has no one in the world really worked on this problem other than in China? Almost 90% of the reference materials come from China. This raises great doubts as to the authors' intentions.

3. It seems to me that using the word equation before every reference to a formula is completely unnecessary. Mdpi standards, from what I remember, require omitting these words.

4. Line 217. Shouldn't it be equation (4) or just (4)? Please check this out.
5. Various letter sizes in equations descriptions and equations are used. This is a bit confusing. e.g. line 217.
5. Please provide more details and more explanations about where the presented calculation formulas came from. On the basis of what conclusions and conclusions were they drawn?
6. Please provide arguments for the presented formulas. Why are the formulas shown in the article are correct? Please connect it with electrical, physical, etc. phenomena.
7. Line 283-284 Please be more specific. I guessed that the numbers in brackets were coordinates, but it's not clear from the article text.
This will allow us to understand the authors' reasoning


Nevertheless, it is advisable to describe the presented formulas and equations in more detail and to link them with physical phenomena or processes. This will allow to understand why the calculations presented in the paper are correct and not obtained e.g. simply by trial and error method. e.t.c.
Apart from the above minor remarks, the article was very professionally written using currently applicable standards.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thanks for the authors' reply. I am satisfied with your revision.

Besides the review round, I want to share my understanding. As mentioned about the labeling, it is really a concern when using machine learning. Actually, I think it is not easy to use the training results in one dataset (such as your training set) in a real scenario. The cross-domain issue and unbalanced sample in real cases may affect the actual performance. If the author can address these concerns and verify in the in-service case, it will be a great contribution.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you to the authors for their response. The explanations provided are sufficient and satisfactory for me.