A Study of Multilayer Perceptron Networks Applied to Classification of Ceramic Insulators Using Ultrasound

Round 1

Reviewer 1 Report

Please follow the standard terms, check if what you mean DIC and FIC instead of DEC and FEC.
Why 2 hidden layers? not 1 or 3 ? 

what library did you use? sklearn?
Define equations 1-4 parameters.
How did you manage the multiple class output of your model only only using sigmoid function? I was expecting softmax!
I could not fathom the comparison you added in section 3, could you state other works that you compared your method to more clearly?
What do you think about the generalizability of NNs models to other instances that are different, e.g. manufacturer, ... etc.

An honest suggestion, I believe the NN model you did is not your biggest contribution in this paper, I think the data set you are using should be a good contribution that can be used as a pivot for this paper. Try to sit and articulate the paper around that data set and the model as a way on how to use it. Of course this suggestion is not binding!!

Author Response

We appreciate for taking the time to read our paper, we forward the answers (in the attached file) looking for the article to be accepted.

Author Response File: Author Response.docx

Reviewer 2 Report

In this paper the authors study the application of Multilayer Perceptron networks to the classification of the different condition of ceramic insulators based on a restricted database of ultrasonic signals recorded in the laboratory.

 

Section 1 must be improved. A paragraph is missing which explains in detail why ultrasound is used to identify defects in insulators. You just say it is a technique for identifying defects, but you don't say how it identifies them. This part is essential otherwise a non-expert reader in the field will not be able to follow the flow of information. It also explains why perforated and contaminated insulators cause power distribution failures. Explain the role of isolators first.

 

Section 2 must be improved. You have to describe in more detail the instrumentation used in the experiments in particular the microphone used for the ultrasound recording. In addition, you have to explain in detail the reason for choosing the descriptors used in the extraction of the features. Finally, you must describe in detail the architecture of the neural network used and the choice of parameters. It seems that a parameter optimization procedure has not been carried out. Finally, but even more importantly, you have to justify in detail the choice of using two neural networks.

 

Section 3 must be improved. The results obtained must be shown more effectively. Figure 5 and 6 can be removed do not add any content to the job. Instead, it would be interesting to add a figure in which the descriptors extracted from the different cases are compared. You also need to spend more time discussing the results obtained with the classification. Explain what the low contaminated / uncontaminated case result may be due to (Table 2). Also discuss in detail all the cases contained in Table 3.

 

Section 4 can be improved. The conclusion lack the possible future goals of this work. Do the authors plan to continue their research on this topic?

 

 

158) NBR 10621/2017? Add reference

169) IEC-507? Add reference

177) ultrasonic microphone. Add the microphone specifications (sensitivity, impedance, Frequency Range, etc.). You can create a table with such values.

187-191) Finally a few lines on the reason for the use of ultrasound. You should move it to the introductory section and add content.

192-196) Then explain how the microphone must be positioned to capture the right components. Add references to papers that have addressed the problem.

209-214) Why did you use these descriptors? Explain this part of the work in detail. The selection of the features is crucial in the elaboration of the input to be presented to the model based on the neural networks.

219) [0.1] ? [0 1] or [0, 1]

217-219) Add a normalization formula

219) Explain in detail how you mixed the samples, or rather how you selected them

226) Introduce the ANNs adequately, they represent the crucial part of the paper. Before talking about the activation function you have to deal with neurons and connection weights. Then you can talk about the essential role of the activation function.

229) Equation 5 lacks the minus sign in the denominator exponent (e^-x)

236) In equation 6 explain the terms. What is yi  and what is ˆ yi

236) Figure 4 is not referred to anywhere in the paper. Also, you should enrich the caption.

244-248) You should insert a flowchart indicating this procedure. This is the weak part of the paper. Why do you have to use two networks? Can you indicate the results you get with 4 outputs? Try to understand the reason for this problem. Have you tried using other descriptors?

248-249)” The number of intermediate neurons is arbitrary and the performance of the network may vary.” In what sense arbitrary? The number of hidden layers and the number of neurons for each layer is essential in the construction of a neural area. It details the architecture of the neural network. Do several tests and carry out a procedure of optimization of the parameters of the network before showing the results.

285) It would have been interesting to compare the features of two different cases

287-293) Then the procedure for choosing the parameters has been carried out. Why didn't you say this in the previous section? Make it clear.

321) Introduce adequately the accuracy

Author Response

We appreciate your review and your suggestions for improving the work. We tried to make the changes and explain what was done in this paper. We forward the answers looking for the article to be accepted in the attached file.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I see you provided clarifications for the comments.

Reading your paper again I recommend you to try to add more insights about the future work.

Author Response

Thank you for your comment. We sent the cover letter with the answers and the revised paper.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors addressed almost all the reviewer's comments with sufficient attention and modified the paper consistently with the suggestions provided. The new version of the paper has improved significantly both in the presentation that is now much more accessible even by a reader not expert in the sector, and in the contents that now appear much more incisive. The addition of a sufficient reference bibliography gave consistency to the authors' statements and the results they achieved.  The detailed description of the figures makes them easier to understand by the author. Finally, the addition of practical applications of the results and the future goals, allows to summarize the meaning of this work.

Unfortunately, some comments were not taken into consideration. For example: you must explain in detail the reason for choosing the descriptors used in the extraction of the features. Why haven't you used any others? You must describe in detail the architecture of the neural network used and the choice of parameters. It seems that a parameter optimization procedure has not been carried out. Finally, but even more importantly, you must justify in detail the choice of using two neural networks. Also you should improve the content of the Figure caption. Add information, remember that the reader should retrieve all information for reading the figure right from the caption. So it must be complete and comprehensive.

 

87) A period in missing at the endo of the paragraph.

266) Remove the period at the endo of the equations

344-344) Do not insert a figure in the text, move it before or after the end of the paragraph.

Author Response

Thank you for your comment. We sent the cover letter with the answers and the revised paper.

Author Response File: Author Response.pdf