Automatic Power Quality Disturbance Diagnosis Based on Residual Denoising Convolutional Auto-Encoder

Round 1

Reviewer 1 Report

1) Line 33- It should be - digital signal 'processing', not process!

2) Literature survey seems a bit weak and adapted for the sake of proposed technique. Authors should also discuss the key challenges in this area of work. What are problems associated with 3 different stages authors have mentioned? Authors should also discuss the need for signal processing in contrast with some recent works on auto-feature selection which do not need signal processing.

3) the time-frequency spectrum of PQDs is divided into 98 two parts, where f ≤ 1.5 f0 and f > 1.5 f0 represent the low-frequency part and high99 frequency part, respectively. - this needs to be elaborated, what is the value 1.5, what is its role, how does it influence spectral analysis? If possible, a graphical explanation showing effect of these separation of time-frequency analysis of pq signal.

4) The values of parameter 100 β1,2 are set to 121 and 10, respectively, according to Ref. [16]. - It is ok to use tested values, but authors should mention in short why these values are well suited for the purpose. 

5)  size of the amplitude time-frequency matrix is 256 × 512 - how come such a small sized matrix for a signal of 2.56 kHz sampling frequency? how many samples will be there in signal? - Authors must present discrete time equations to answer these questions!

6) please refrain from using the word handcrafted features. this is a bad practice. there is no craftwork involved. it is known as feature engineering for a reason. manually selected features or manual feature selection is more appropriate.

7) Given an input amplitude time-frequency data set x = {x1, x2, ..., xl} with shape size m × n for each element - x is one dimensional, them how the shape size became two dimensional? where did m, n come from? what are those parameters?

8) Before talking about the mapping, I think the approach and purpose of mapping should be explained. Maybe a figure should have been included so that a visual understanding is also possible

9) There is no mention of what were the parameters chosen for the cnn. what values and how were they chosen? what is the learning algo used? what is the learning rate & dropout rate?

10) Based on visual inspection Layer 1 seems to represent L13 PQD better as compared to Layer 3. Comment on that! On what basis, do you choose the number of layers? How do you judge its efficacy?

11) If the authors decided to show such small dimension / resolution boxes for spectral content of PQDs, then atleast they should have present more cases with subfigures in Fig. 3. 

12) Which harmonic levels are present in L5?

13) Authors should have shown some input signals, because fig. 4 L9 doesn't show any presence of spike? What is a voltage spike how does it occur?

14) Fig 8 title - 'with varying noise'!

15) why did authors decide to compare RDCA, CAE, and CNN? Is it a benchmark?

16) Only 50 samples per PQD is not sufficient for verification. It is not possible to have too many variations for training the classifier. how can you be sure it is not baised? And testing time of >200ms if actually quite high to just test 50 signals. Obviously the CNN is taking too much time for processing. What value does the proposed method have in that case?

 

 

Author Response

   Thank you very much for your careful comments. We have revised the manuscript carefully and given a detailed response to each comment. The specific response file has been attached.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see the attached pdf file.

Comments for author File: Comments.pdf

Author Response

Thank you very much for your careful comments. We have revised the manuscript carefully and given a detailed response to each comment. The specific response file has been attached.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Still some concerns remain:

1) Please include all equations (esp. for digital algorithms) in the revised manuscript. It is not only for the reviewers, but more important for the final readers to understand and implement what you have done.

2) Please include some of those enlarged images and explanations of spectral analysis in manuscript. 

3) Also, show spectral analysis of oscillatory transients detection to demonstrate the ability of your technique.

4) In the introduction there is a line - However, unsuitable parameter
setting will impact the time-frequency analysis performance. - Why would anyone use unsuitable parameters for analysis? It is like saying - those people were wrong because they would use fork to have some soup - why would they ? Is that the only criticism you have got for works in literature? 

Author Response

Thank you very much for your careful comments. We have revised the manuscript carefully and given a detailed response to each comment. The response file has been attached.

Author Response File: Author Response.pdf