Safety Risk Assessment Using a BP Neural Network of High Cutting Slope Construction in High-Speed Railway
Round 1
Reviewer 1 Report
This paper is using an ANN model for examining Safety Risk Assessment of High Cutting Slope Construction in High-speed Railway. The idea of this study is very interesting and contributing, however the methodology and development part related with the ANN model, has many major mathematical issues.
The biggest mathematical problem is related with the overfitting phenomenon, since the authors are using a small data set (n=48) and the number of ANN ‘s input parameters equals with 39. This number of input parameters is un-proportionally very big, therefore the created ANN can’t be reliable, because is affected by the overfitting phenomenon.
Additionally, the ANN’s methodological part is not properly/clearly described. The writing style is very confusing and it’s hard to understand trivial/ well known theory (e.g., wrong terms are used). For example:
- Line 171: “only one set of data” : the word “data”is wrong, it should be parameter.
- Line 189-190: “ In order to avoid too much data causing the neural to be less accurate and the difference value between the input value and the threshold value to be too large.” This is not true statement, actually ANNs are data starving. Probably the word “data’ should be parameters.
- Line 218: instead of “traindx’ should be “traingdx”.
Author Response
Dear reviewer, thank you for your comments! For your comments, we have made the corresponding responses. Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
I think this is written in hurry. This is very clear from figure preparation, Text writing, methodology. Paper is hard to follow. The paper itself is nice but need more experiment and organization. I have the following main points:
1- Quality of the figures need to be improved
2- Organization of the paper should be modified to avoid the confusion. Section 2 and 3 should be merged in only one section "Data and methods"
3- Case study section should be moved to results section
4- Most of that table should be moved to appendix. The reader became distracted by those too much tables (from table 9 to 22)
5- Table 20 Error analysis is shown after table 24. There are two tables with number 20, in page 11 and page 14
6- Rather than that standard figure 1 which is generated from Matlab, please show your own figure that contain name of the inputs and outputs for your NN.
7- Many evaluation parameters could be used and compared. since every parameter provide different perspective about the trained model.
8- If the source code that used for NN. model will be available for community, You should put it on any open rep. (e.g. gitlab, github,...etc.)
9- Abstract and summary section should be rewritten in easy reading way to the reader
10- More test cases could be used if possible
Author Response
Dear reviewer, thank you for your comments! For your comments, we have made the corresponding responses. Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments
Thank you for taking my comments into consideration. As the authors clarify/state in their revised version the data samples number= 216. Therefore, my main consideration regarding the overfitting phenomenon cannot exist anymore, based on this authors’ statement. However, I believe the following modifications would be beneficial for this manuscript.
- Please provide some more information about Figure 5 (new line 310) and why you don’t use the whole data set for this simulation (the data points number are clearly much less than 216)? Additionally Figures 3 & 4 I don’t really believe that provide any useful information and I would suggest the authors to remove them.
- Regarding Table 9 (new line 316), I believe that a graphical representation would be more elegant. Please, plot the real vs the predicted data for your data set.
- New line 312-313: “The error analysis of the expected output and the actual output shows that the maximum error is only 0.995%” the ANN is evaluated based on the test set error. Please revise accordingly
- New line 293-294: please provide a graphical representation of these weight values for each parameter
- New lines 164-165: please replace the word “data”with the word “parameter”. Adding a relevant reference is recommended.
- New lines 166-167: “Meanwhile, And the selected data should take into account the characteristics of a high cutting slope in HSR, and if a sample with a large error occurs, it can not be selected”. Please remove these part
- New lines 176-177: As mentioned before, the output nodes in this study are 39 and the input nodes are 1” please remove this part
Author Response
Thank you for your comments! For your comments, we have made the corresponding responses. Please see the attachment.
Author Response File: 
Author Response.docx
Round 3
Reviewer 1 Report
Thank you for taking my comments into consideration and try to make the suggested modifications. The revised version of the manuscript is significantly improved; however some minor issues exist.
- Lines 173-174: “Therefore, the sample selection should meet both quality and quantity to obtain a reliable BP network with accurate prediction ability”. Please remove this part, it is not necessary.
- Lines 174-176: “In this study, the sample selected in this paper is the high cutting slope and the samples were obtained by questionnaire survey”. These lines are not clear. Do you mean the parameters instead of sample? Please revise this.
- Lines 282-283: “The results of the error analysis for the first 48 sample data are shown in Figure 5.” Why do you use for Figure 5 the 48 first sample data? It is recommended to use only the data of the test set to study the error and create the Figure 5. Please adjust the text and Figure 5 accordingly.
- You mention Matlab in Lines 214 and 271. Please mention that you use Matlab only in Line 214.
- Please check the format of your references, especially the newly added # 47-49.
Author Response
Thank you so much for your comments! The authors have made the corresponding responses. Please see the attachment.
Author Response File: Author Response.docx
