Ultrasonic Assessment of Thickness and Bonding Quality of Coating Layer Based on Short-Time Fourier Transform and Convolutional Neural Networks

Round 1

Reviewer 1 Report

My humble opinion is that paper is very interesting, however several points must be discussed and amended:

• Title Condition Monitoring of Highly…what is Title condition, a mistake or a new approach?
• I do not see why the continuous monitoring is so important, perhaps doing it time to time could be enough
• At states and Europe the conclusions are usually given as point and giving quantification is much better. Things are not good…thing must be %percentage good.
• ALOCIT 28.14 and RS 500P does they have other names, more standard. Just in the contrary, explain from which machines or technologies are obtained.
• In some sources such as https://doi.org/10.1007/s11665-016-2343-6 US are defined as basic approach but several modifications must be taken into consideration. Here the control is key.
• The Keras library was used in the Python programming environment to build the proposed CNN model….this phrase is from your text, perhaps you can reduce a little, or please explain better. Python is important here?
• I see Figure 12, OK. Some authors also proposed Smart optimization process based on boosting ensembles, Journal of manufacturing systems. Did you check this journal?
• Again, Two different types of coating materials were used in the experiments: RS 500P (Figure 2(a)) and ALOCIT 28.14. pleas define them, US will depend on the layers, thicknesses and so on. How many repetitions did you do of testpieces?
• You explain well the fundamentals, US is very typical for layer thickness evaluation.
• Did you check works by Ciarrocchi, or Celaya, or Mayer, or Aurrekoetxea in MDPI or others…the authors worked on US or residual streeses and I can not see them.

Making a final summary:

Format of the paper is good and some figures are OK, those about results. A better discussion and more literature review is expected. It is the only way to prevent from require a full testing campaign of mechanical performance such as fatigue, creep, scratch testing… define better which are the coatings.

Major can be a logical decision, but paper could be easily enhanced, I beg you to work in this lines.

Author Response

Response to Reviewer 1  

Thank You for your kind comments, we appreciate all of them. According to your comments, we tried to fix and add details to our manuscript. Details of the corrections are following: 

Title Condition Monitoring of Highly…what is a Title condition, a mistake or a new approach?

Thank you for your notice. The mistype was corrected.

I do not see why the continuous monitoring is so important, perhaps doing it time to time could be enough

The title has been changed considering the comments of the reviewer to:  Ultrasonic assessment of coating layer thickness and bonding quality based on the Short Time Fourier Transform and Convolutional Neural Networks    

At states and Europe the conclusions are usually given as point and giving quantification is much better. Things are not good…thing must be %percentage good.

The conclusion of the manuscript was rewritten, considering the comments of reviewer (L549-571)

ALOCIT 28.14 and RS 500P does they have other names, more standard. Just in the contrary, explain from which machines or technologies are obtained.

Additional information was added: RS 500P (Chemco, Coatbridge, Scotland, UK) and ALOCIT 28.14 A&E Group, Shah Alam, Malaysia) were used as coating materials in the experiments.  And both coating materials are applied by conventional roller. (L289 -L294).  Also, a reference related to the coating materials is added.

"Ref: Won, B., Kim, M. O., Park, S., Yi, J.-H., Effects of Water Exposure on the Interfacial Bond between an Epoxy Resin Coating and a Concrete Substrate, Materials 2019, 12, 3715; doi:10.3390/ma12223715"

In some sources such as https://doi.org/10.1007/s11665-016-2343-6 US are defined as basic approach but several modifications must be taken into consideration. Here the control is key.

The suggested paper describes interesting researches regarding machining and ultrasonic vibration. However, our research scope is different, differently from the vibration we deal with wave propagation.

The Keras library was used in the Python programming environment to build the proposed CNN model….this phrase is from your text, perhaps you can reduce a little, or please explain better. Python is important here?

We appreciate your comments. Yes, the Keras library can be also be implemented in other programming environments, for example, Java so on. Correspondingly, it was shortened.

I see Figure 12, OK. Some authors also proposed Smart optimization process based on boosting ensembles, Journal of manufacturing systems. Did you check this journal?

Considering your comment additional information was added regarding to the parameters of the CNN. The main task was to implement the CNN, and in the further researchers, we are going to optimize the CNN parameters. Additional statements was added in lines L261-278

Again, Two different types of coating materials were used in the experiments: RS 500P (Figure 2(a)) and ALOCIT 28.14. pleas define them, US will depend on the layers, thicknesses and so on. How many repetitions did you do of testpieces?

Mistype was corrected. Additional information corresponding to the difference between measurements errors was added as a table. Please refer to Table3 on page 11

You explain well the fundamentals, US is very typical for layer thickness evaluation.

Thank you for your positive comments, one of the main difficulties of ultrasonic measurement of the coating layer is the high attenuation energy and thin thickness of the layer.

Did you check works by Ciarrocchi, or Celaya, or Mayer, or Aurrekoetxea in MDPI or others…the authors worked on US or residual streeses and I can not see them.

 The suggested authors did interesting researches regarding machining with ultrasonic vibration. However, our research scope is different from them. We use US wave propagation. A literature has been added https://doi.org/10.1016/j.ndteint.2018.04.002 regarding US-based bonding evaluation.

Author Response File: Author Response.pdf

Reviewer 2 Report

The article investigates the ultrasonic inspection of the coating materials. Firstly, it uses the magnitude of the STFT to determine the bonding condition and thickness of the coating layer. After, a CNN combined with the STFT is used in the automatic detection of the coating layer condition. The work presents interesting results.

Section 2.1, the STFT method must be better described and explained.

Section 3.1, provide an error analyses in the estimation of the thickness of the two coating materials when compared with the real thickness measured by the accurate “Mitutoyo” calipers.

The activation function ReLu is not well defined by equation (9). It should be: 0, if x<0; x, if x>=0.

In section 4, the authors should provide more details about the CNN, such as the learning algorithm and hyperparameters used during the training of the neural network, and so on.

Add more comments about the importance of the CNN for the problem under study. Clearly state the advantages and drawbacks of this technique.

We detect some minor misprints. Read carefully the manuscript and correct them.

Author Response

Response to reviewer 2

Thank you for your kind comments on our manuscript. We appreciate your comments. We tried to modify and to revise according to your suggestion. Details of modification are following:

Section 2.1, the STFT method must be better described and explained.

Definition and description of STFT has been added. Details regarding the windowing function and 3D plots were included, also was briefly compared with simple FT (Fourier Transform).

Section 3.1, provide an error analyses in the estimation of the thickness of the two coating materials when compared with the real thickness measured by the accurate “Mitutoyo” callipers.

An additional table (page 11 Table 4) was added, considering errors in the comparison between calipers and ultrasonic measurement.

The activation function ReLu is not well defined by equation (9). It should be: 0, if x<0; x, if x>=0.

Typo was corrected, and an additional description of ReLU activation function was included.  (L242-245)

In section 4, the authors should provide more details about the CNN, such as the learning algorithm and hyperparameters used during the training of the neural network, and so on.

Additional information was added regarding the details of the CNN. (Table 1 and Table 2) on page 19, also description and details of applied hyperparameters were stated.  

Add more comments about the importance of the CNN for the problem under study. Clearly state the advantages and drawbacks of this technique.

Additional information corresponding to the pros and cons of the CNN was included (L270 -276)

We detect some minor misprints. Read carefully the manuscript and correct them.

Thank you for your comments, the misprints were checked.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I am afraid that author did not attend the suggestions and comments. Only tipos were corrected. The discussion, literature review and general paper text is more or les the same.

Please improve following the points or just withdraw the work, at States and Europe papers under Major must be very improved.

Author Response

Response to Reviewer 1  

Thank You for your kind comments, we appreciate all of them. According to your comments, we tried to fix and add details to our manuscript. Details of the corrections are following:   

1.Title Condition Monitoring of Highly…what is a Title condition, a mistake or a new approach?

Extra word “Title” was removed. Also, the original title: “Condition monitoring of highly attenuative coatings based on short-time Fourier Transform and Convolutional Neural Networks” was modified to “Ultrasonic assessment of thickness and bonding quality of coating layer based on Short Time Fourier transform and Convolutional Neural Networks” The title was modified to underline used methods, also the word “monitoring” was not relevant.

2. I do not see why the continuous monitoring is so important, perhaps doing it time to time could be enough

We agree with your comment, also the title of the manuscript was modified to avoid misunderstanding (Please refer to comment 1). Also, the sentence was used “The thickness and adhesion state are two important features of coatings, which must be monitored periodically (L37-38)”.

3. At states and Europe the conclusions are usually given as point and giving quantification is much better. Things are not good…thing must be %percentage good.

The whole contents of the manuscript conclusion have been rewritten by adding the details of results (please refer to lines: L549-571).

Here is a brief summary of conclusions:

Comparing the thickness measurement results of precise calipers and ultrasonic measurements showed slight difference. The difference was about 1% for both types of coating material.

Detection of the debonded coating layer based on the spectrogram of the waveform and CNN is feasible and effective. Applied CNN-based approach showed over 99 % of correct classification of the bonded and debonded condition of the coating layer.

4. ALOCIT 28.14 and RS 500P does they have other names, more standard. Just in the contrary, explain from which machines or technologies are obtained.

1)

28. a) ALOCIT 28.14 is an official product name of the coating material from A&E Group. This was referred as Epoxy A in the manuscript. (Please, refer to Table 1 on page 5)

1. b) RS 500P is an official product name of the coating material from Chemco. In the manuscript it was referred as Epoxy B. (Please, refer to Table 1 on page 5)

2)

Both coating materials are applied by conventional roller. Also, a reference related to the coating materials was added.

"Ref: Won, B., Kim, M. O., Park, S., Yi, J.-H., Effects of Water Exposure on the Interfacial Bond between an Epoxy Resin Coating and a Concrete Substrate, Materials 2019, 12, 3715; doi:10.3390/ma12223715"

5. In some sources such as https://doi.org/10.1007/s11665-016-2343-6 US are defined as basic approach but several modifications must be taken into consideration. Here the control is key.

The proposed paper describes an interesting study on machining and ultrasonic vibration. And like you said, control is very important in the case of high power ultrasound. However, the scope of our research is different, as it deals with wave propagation..

6. The Keras library was used in the Python programming environment to build the proposed CNN model….this phrase is from your text, perhaps you can reduce it a little, or please explain better. Python is important here?

We humbly accept your opinion. Keras is an open-source library, which can use in Python, Java and other programs. To provide more details used program we mentioned Python. The sentence was shortened.

7. I see Figure 12, OK. Some authors also proposed Smart optimization process based on boosting ensembles, Journal of manufacturing systems. Did you check this journal?

In the suggested research (https://doi.org/10.1016/j.jmsy.2018.06.004), the author used AdaBoost algorithm to find optimal parameters of neural networks. This method is efficient for optimize the parameters of neural networks. However, in our research we did not concentrate our attention on parameters optimization of CNN, we only wanted to show the performance of simple CNN in the classification of bonding condition of coating. AdaBoost and other optimization methods are going to be applied in our future researches. Following your comments, additional infromation about types of neural networks parameters boosting algorithms were added.

Also, following references were added:

1. 1. Weldegebriel, H.T.; Liu, H.; Haq, A.U.; Bugingo, E.; Zhang, D. A New Hybrid Convolutional Neural Network and EXtreme Gradient Boosting Classifier for Recognizing Handwritten Ethiopian Characters. IEEE Access 2020, 8, 17804–17818, doi:10.1109/ACCESS.2019.2960161.
2. Taherkhani, A.; Cosma, G.; McGinnity, T.M. AdaBoost-CNN: An Adaptive Boosting Algorithm for Convolutional Neural Networks to Classify Multi-Class Imbalanced Datasets Using Transfer Learning. Neurocomputing 2020, 404, 351–366, doi:10.1016/j.neucom.2020.03.064.
3. Thongsuwan, S.; Jaiyen, S.; Padcharoen, A.; Agarwal, P. ConvXGB: A New Deep Learning Model for Classification Problems Based on CNN and XGBoost. Nuclear Engineering and Technology 2021, 53, 522–531, doi:10.1016/j.net.2020.04.008.
4. Sudakov, O.; Burnaev, E.; Koroteev, D. Driving Digital Rock towards Machine Learning: Predicting Permeability with Gradient Boosting and Deep Neural Networks. Computers & Geosciences 2019, 127, 91–98, doi:10.1016/j.cageo.2019.02.002.
5. Bustillo, A.; Urbikain, G.; Perez, J.M.; Pereira, O.M.; Lopez de Lacalle, L.N. Smart Optimization of a Friction-Drilling Process Based on Boosting Ensembles. Journal of Manufacturing Systems 2018, 48, 108–121, doi:10.1016/j.jmsy.2018.06.004.

8. Again, Two different types of coating materials were used in the experiments: RS 500P (Figure 2(a)) and ALOCIT 28.14. pleas define them, US will depend on the layers, thicknesses and so on. How many repetitions did you do of testpieces?

The thickness of the specimens was measured 5 times repeatably by caliper and also 5 times measurements were performed by ultrasonic testing. Corresponding average values and standard deviations are added in Table 2.

Table 2. Coating layer thickness measurements by caliper and ultrasound 

9. You explain well the fundamentals, US is very typical for layer thickness evaluation.

The thickness evaluation by US is typical as you mentioned, however that is true for low attenuative materials. The coating materials concerned this research show very high attenuation, so that extensive research is necessary.

10. Did you check works by Ciarrocchi, or Celaya, or Mayer, or Aurrekoetxea in MDPI or others…the authors worked on US or residual streeses and I can not see them.

The suggested authors did interesting researches. Most of the authors like Amenabar et al, Mayer et al, Mor et al. did research on the correlation between quality and defect adhesion of joints with ultrasonic using bulk waves parameters. Also, He et al. applied a guided wave for lap joint defect detection. Mayer et. al used reflected echo phase for ultrasonic Imaging technique and materials classify defect. Considering their contributions following references were include in the manuscript.

1. Celaya, A.; Lacalle, L.N.L. de; Campa, F.J.; Lamikiz, A. Ultrasonic Assisted Turning of Mild Steels. IJMPT 2010, 37, 60, doi:10.1504/IJMPT.2010.029459.
2. Aurrekoetxea, M.; López de Lacalle, L.N.; Llanos, I. Machining Stresses and Initial Geometry on Bulk Residual Stresses Characterization by On-Machine Layer Removal. Materials 2020, 13, 1445, doi:10.3390/ma13061445.
3. Fitzka, M.; Mayer, H. Variable Amplitude Testing of 2024-T351 Aluminum Alloy Using Ultrasonic and Servo-Hydraulic Fatigue Testing Equipment. Procedia Engineering 2015, 101, 169–176, doi:10.1016/j.proeng.2015.02.022.
4. Ozen, F.S.; Celaya, M.; Nazarian, S.; Saltan, M. J I T OURNAL OF NNOVATIVE RANSPORTATION. 9.
5. Amenabar, I.; Mendikute, A.; López-Arraiza, A.; Lizaranzu, M.; Aurrekoetxea, J. Comparison and Analysis of Non-Destructive Testing Techniques Suitable for Delamination Inspection in Wind Turbine Blades. Composites Part B: Engineering 2011, 42, 1298–1305, doi:10.1016/j.compositesb.2011.01.025.
6. He, J.; Guan, X.; Peng, T.; Liu, Y.; Saxena, A.; Celaya, J.; Goebel, K. A Multi-Feature Integration Method for Fatigue Crack Detection and Crack Length Estimation in Riveted Lap Joints Using Lamb Waves. Smart Mater. Struct. 2013, 22, 105007, doi:10.1088/0964-1726/22/10/105007.
7. Mayer, K.; Langenberg, K.-J.; Krause, M.; Milmann, B.; Mielentz, F. Characterization of Reflector Types by Phase-Sensitive Ultrasonic Data Processing and Imaging. J Nondestruct Eval 2008, 27, 35–45, doi:10.1007/s10921-008-0035-3.
8. Mayer, K.; Ibrahim, M.; Krause, M.; Schubert, M.; Kassel, U. Requirements for a Small Size Ultrasonic Imaging System for Inspection of Concrete Elements. 9.
9. Mayer, K.; Chinta, P.K.; Langenberg, K.-J.; Krause, M. Ultrasonic Imaging of Defects in Known Anisotropic and Inhomogeneous Structures with Fast Synthetic Aperture Methods. 10.
10. Norambuena-Contreras, J.; Castro-Fresno, D.; Vega-Zamanillo, A.; Celaya, M.; Lombillo-Vozmediano, I. Dynamic Modulus of Asphalt Mixture by Ultrasonic Direct Test. NDT & E International 2010, 43, 629–634, doi:10.1016/j.ndteint.2010.06.007.
11. Mor, E.; Aladjem, M.; Azoulay, A. A Sparse Approximation Method for Ultrasonic Monitoring the Degradation of Adhesive Joints. NDT & E International 2018, 98, 17–26, doi:10.1016/j.ndteint.2018.04.002.

Author Response File: Author Response.pdf

Round 3

Reviewer 1 Report

Good changes