Deep Machine Learning for Acoustic Inspection of Metallic Medium

Round 1

Reviewer 1 Report

Thanks to the authors for performing this research on "Deep Learning For Materials Inspection". The paper is interesting and reads well. All topics are carefully explained and allow easy understanding even for a reader unfamiliar with the subject. The following comments need to be addressed:

(Q) In Fig.5,8,10,11...,is difficult to identify because the letters and labels are small. Edit all the Figure in paper in high-definition.

Author Response

In Fig.5,8,10,11...,is difficult to identify because the letters and labels are small. Edit all the Figure in paper in high-definition.

 

I will work on improving the clarity of these figures. Thank you for your review.

Reviewer 2 Report

This paper shows a very interesting analysis about potentialities given by deep learning tools, such as convolutional neural network in applications concerning with material inspection and damage detection. In particular, the authors mange to apply a neural classifier on a dataset obtained by analyzing the wave propagation on a steel ring. The paper is well written and results are interesting also in consideration of potential perspective of applications on real defects. To improve the paper I would suggest to enrich the state of the art part in the Motivation section. There is some recent research in adopting CNN and k-fold approaches for acoustic signal and acoustic emission I think that giving more evidence of them would help in making clear the originality of the work.

In the following I am adding some more comments:

 

Pag. 3 Lines 131-133

What is reported in these lines regarding amplitude reduction and phase shift is correct but authors refer to the reflected part while transmitted signal is observed. Could the author clarify this passage?

Pag.5 Materials and Methods

I think it would be useful to ad a picture of the experimental set-up.

Pag.5 Lines 194-195

Is 400 kHz the acquisition frequency of the oscilloscope? By looking at line 202 it seems to be the frequency of the emitted frequency. By Nyquist the frequency for collecting data should be at least twice. Could you clarify this point?

Pag.8 Background

I would recommend to reference this general part on ANN, it would be easier for the reader to have some points where to go deeper in these concepts

Pag.10 Line 346 Input

Can the author add an example of one of  200x15x1 elaborated? I think it would be a more direct visualization of the structure of the input data.

Pag.11 Frequency domain signals. I do not understand well the title of this section in consideration of the fact that the core of the sections is more connected with the spatial representation rather than frequency representation

Pag.13 Results on CNN

Can the author specify the size of the training and validation set?

Pag.14

Figures 12 and 13 are hardly visible at 100% I would suggest to provide a more readable version

Author Response

What is reported in these lines regarding amplitude reduction and phase shift is correct but authors refer to the reflected part while transmitted signal is observed. Could the author clarify this passage?

This sentence aims to explain why the phase and amplitude can be examined for the presence of an anomaly. Because a portion of the signal will be reflected back in presence of the anomaly, it will change how the transmitted signal is observed. Perhaps this is better phrased as: “At the interface with the anomaly, the acoustic wave experiences partial reflection that causes
reduction in amplitude and phase shift in the transmitted signal.”

I think it would be useful to ad a picture of the experimental set-up.

I agree. We will add an image of the experimental setup.

Is 400 kHz the acquisition frequency of the oscilloscope? By looking at line 202 it seems to be the frequency of the emitted frequency. By Nyquist the frequency for collecting data should be at least twice. Could you clarify this point?

400 kHz is the frequency of the signal being emitted by the transmitter of the oscilloscope. I will update this sentence to clarify.

I would recommend to reference this general part on ANN, it would be easier for the reader to have some points where to go deeper in these concepts

I will add a little more information on the components of ANN with a reference on ANNs.   

Can the author add an example of one of  200x15x1 elaborated? I think it would be a more direct visualization of the structure of the input data.

I will add an image to help illustrate this.

Frequency domain signals. I do not understand well the title of this section in consideration of the fact that the core of the sections is more connected with the spatial representation rather than frequency representation

This is a good point, thank you for catching this. This sub-section title is not quite fitting for the information in the section. We will rename this section to “Modal Analysis” as that is more in line with the contents of this section.

Can the author specify the size of the training and validation set?

We allocated 315 test images and 315 training images. In a previous section, we referred to this as a 50-50 split, but I will add this detail to this section of the paper so that it is readily available to the reader.

Figures 12 and 13 are hardly visible at 100% I would suggest to provide a more readable version

I have resized these images to make them more clear.

Reviewer 3 Report

The authors proposed a deep learning model to identify structures exhibiting signs of damages. Deep learning is a great tool for engineering and science, which is a great topic. Here are some suggestions for the authors. 

1. The title is too big, I would suggest the authors to write a more specific title for this paper.

2. The introduction part needs a deep review on the previous researches, more papers need to be cited, here is just one example showing machine learning application in acoustic emission, but you need more than that.

Li, D., Zhang, S., Yang, W., & Zhang, W. (2014). Corrosion monitoring and evaluation of reinforced concrete structures utilizing the ultrasonic guided wave technique. International Journal of Distributed Sensor Networks, 10(2), 827130.

3. In lines 257 to 267, I suggest you use a table to show all these labels.

Author Response

1. The title is too big, I would suggest the authors to write a more specific title for this paper.

We will re-title this paper to "Deep Machine Learning for Acoustic Inspection of Metallic Medium"

2. The introduction part needs a deep review on the previous researches, more papers need to be cited, here is just one example showing machine learning application in acoustic emission, but you need more than that.

We have added reference to some of the other papers we have reviewed.

3. In lines 257 to 267, I suggest you use a table to show all these labels.

We have moved the label definitions to a table

Round 2

Reviewer 3 Report

All the comments are well addressed. 

Author Response

Thank you for your review!