Evaluation of Internal Cracks in Turbine Blade Thermal Barrier Coating Using Enhanced Multi-Scale Faster R-CNN Model

Round 1

Reviewer 1 Report

Dear authors,

With interest I have read your manuscript about the crack evaluation in thermal barrier coatings. From the user perspective, it would be a great progress, if crack detection or delamintation in TBCs could be clearly detected after coating deposition or maybe as well during service intervals of turbines. Below I summarize by comments and necessary improvements:

• The manuscript contains many abbreviations, which are only sometimes introduced in the text. Hence, a list of abbrevitions should be included in the manuscript as well as the introduction of all abbreviations should be checked in the text.
• Upper and lower case letters are partly used incorrectly. Check the whole manuscript.

• Abstract: The main findings of your study should be outlined in the abstract, compare with the conclusion.
• 1. Introduction (line 88 - 95): Revise the scope of the work and describe it according the abstract and the manuscript in more detail. I suggest to include the applied methods including the abbreviations. Thus, it is easier for the reader to follow the manuscripts structure.
• Figure 2 in line 273 should be moved to the relevant text passage in line 114.
• 3.1 Dataset description, line 244-245: The coating properties, porosity, crack width and length, should be specified and the chosen values explained in more detail. It would be interesting to include, which properties, and their values, are really relevant for TBCs.
• 3.3 Experimental Setup, line 261: Include information about the TensorFlow and ADAM software.
• 4.1 Evaluation criteria, Equation (12): For me the 2 is not clear in the equation.
• Figure 3 and Figure 4: The images look almost identical. Please comment, where are they come from? What are the signs in the figures that indicate a crack? Furthermore, include a unit for the colour scale.
• 5 Conclusions, line 481 - 482: What do you mean by "under different coating lengths"? Please revise.
• 5 Conclusions, line 483 - 487: The further research approach should be explained in more detail in comparison to the presented study, e.g. "other charasteristic cracks", etc.

• Line 208: "..., it is difficult for me to determine,..." Revise using scientific English language.

After revision the manuscript can be published. All the best.

Author Response

Please see the attachment，thank you.

Author Response File: Author Response.docx

Reviewer 2 Report

Authors chose the current interest topic of interest to many sectors. However, it looks like a project report instead of a scientific article; if authors could able to refine the structure will become more reader friendly. 

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

Please find my comments in the attached file.

Best regards

Comments for author File: Comments.pdf

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Reviewer 4 Report

Interesting article, showing another model of discontinuity detection in TBC coatings. TBC coatings are used as most often ceramic coatings to limit the effects of high temperature on a metal material (substrate). The use of the coating should extend the service life of the product. Therefore, it is important to perform periodic inspections and possible repairs of coatings.

On the other hand, the cracks or porosity referred to in the article arise naturally during the application of the coating or during operation due to the dimensional change of the substrate during cooling and heating (thermal expansion). Especially the operational ones are discontinuities full of products from the process.

Regarding the presented algorithm itself, it is presented clearly and correctly. Not less to consider such comments: Thus, the conditions of formation and the morphology of discontinuities (cracks or porosity) are influenced by the type of materials used, both the TBC and the substrate, and in particular their physical properties at room and elevated temperature. These parameters have not been defined in the article (this applies not only to this publication but also to many others).

Moreover, straight cracks were selected for validation analysis, while they are often non-straight and curved or branched in nature. Cracks are also accompanied by chipping on the sides. Here are the questions: What is the detection rate for this type of crack? Were they so analyzed? What is the mismatch / error ratio for this type of crack?

SEM images were used for the analysis of microcracks, while in assessing the condition of the coating under operating conditions, such tests are performed as an effect. Or maybe for microcracks it is worth using high-resolution images made with conventional techniques, and SEM only to verify the correctness of the algorithm and detectability?

What is the practical use of the algorithm? Is it suit able for liners in power plant?

The TBC’s surface pollution influence on the results of analysis. If yes, what kind of polution has stronger influence metallic or non-metallic?  

An important element in the discussion is the inability to clearly distinguish between porosity and small cracks. Why is this not in the conclusions?

Due to technical remarks - pictures 3 and 5 are illegible.

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Reviewer 5 Report

Re.:            Applied Sciences, manuscript applsci-1627811

Title:           A Multi Scale Enhanced Faster R-CNN for Internal Crack Evaluation in Turbine Blade Thermal Barrier Coating

Authors:      Licheng Shi, Yun Long, Yuzhang Wang, Xiaohu Chen and Qunfei Zhao

 General Statement

In the paper a novel image processing methodology for infrared nondestructive inspection of thermal barrier coatings (TBSs) has been presented. The methodology is based on a big data science neural algorithms i.e. convolutional neural networks (CNN). The classification training was performed on an isomorphically multiplied literature data. The algorithm performance was illustrated by comparison with results obtained with some other popular image recognition algorithms. The results were complemented with additional analysis i.e. prediction of the life time TBC performance.

In spite of the fact that the paper presents only simulated data it comply with the Applied Sciences journal profile. With no any doubt the presented issue is up to date also. However, due to some deficiencies in the metric presented, I cannot recommend publication as it stands. The major problem is that it is almost impossible to assimilate the content and convince yourself of the correctness of the work without a thorough reading of the cited literature. It mostly concerns the problem core i.e. the NDT data source and representativeness. The article is too heavily biased toward CNN tech, and such material is better suited to IT journals. In order to be recommended to the Applied Sciences it should be or shortened to a just technical note with presentation of the algorithms comparative study or complemented with some more detailed information about the TBC source data.

Personally, I am convinced of the correctness of the work, but I have gained this conviction indirectly. To be sure, I recommend making additions according to the comments below.

With regard to the current manuscript state, I recommend publication after major revisions.

Major Comments

1.    I will make the most important remark with reference to the details of the Fig. 3 by explaining the problem in further reasoning. If the presented thermograms (simulated thermograms) are images of a stationary state, it is hard to believe that from them it is possible to identify defects as it is shown in the lower parts of each frame. Because of that I presume that not only single frames were analyzed but sequences of them showing a TBC system response to a certain thermal excitation. If yes, the details concerning the “experiment” have to be provided in the paper. Readers should know what type of the NDT procedure is analyzed. And the illustrative material needs to be complemented with a more detailed Fig. 3 single case illustration.

2.    if not, that is, if the data come from stationary “experiments”, I also recommend presenting intermediate data from a single-case analysis.

3.    The second my observation concerns the perfunctory treatment of the data classification, training and evaluation problem, in light of the very precise description of parts of the algorithm (see lines 132 – 232). The authors should provide information about the proportions between the generated data used for the classification and network training and for the procedure evaluation. Do the same data with the same proportions were applied while performing analyses with alternative procedures?

4.    The last point seems minor, but the very large number of unexplained abbreviations and jargon names for those interested in TBC technologies but less familiar with image recognition makes the work impossible to understand. I have tried to highlight this in the text using pop-up notes (more than 20 of them – see the attached pdf file).

Minor Comments

5.    Please check the paper for a possible editorial mistakes (like in line 526)

6.    Did any kind of sensitivity analysis have been performed (e.g. while processing “disturbed” images)?

Comments for author File: Comments.pdf

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Reviewer 6 Report

The document needs revision to improve clarity of presentation, discussion of experimental methods and results.

Figures should be readable also in legend and alberl, and this is not the case

Research methods should be properly introduced, the reader should be potentialli able to reproduce the same study.

In the present form, the manuscript gives no enough control to the reader

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Authors addressed my comments and the manuscript may please be accepted in the present format.

Author Response

We would like to thank the Reviewer again for taking the time to review our manuscript and give helpful comments. We are very appreciate for the Reviewer’s affirmation of this manuscript. 

Author Response File: Author Response.docx

Reviewer 5 Report

Re.:              Applied Sciences, manuscript applsci-1627811_ver2

Title:             Evaluation of Internal Cracks in Turbine Blade Thermal Barrier Coating Using Enhanced Multi-Scale Faster R-CNN model

Authors:      Licheng Shi, Yun Long, Yuzhang Wang, Xiaohu Chen and Qunfei Zhao

 General Statement

I accept all the introduced changes. The authors' explanation cleared up most of my doubts. One other point remains to be clarified, perhaps less significant but important in my opinion for understanding the work. Weighing the decision between minor and major corrections, I lean toward the latter. This is because, through the authors' response, I want to gain confidence in the correct understanding of the content of the paper.

With regard to the above mentioned, I recommend publication after major revisions.

Detailed Comment

While responding:

“Response 2:Thanks to the Reviewer’s very deep comments. Our numerical coating model is based on the assumption that the pore distribution of the coating remain unchanged in a period of time (This is normal in reality, because the pore location of the coating are determined at the time of manufacture. Only some extreme cases, like corrosion, can change the pores). Therefore, the generated dataset can be regarded as an ‘image sequence’ to some extent(since crack grows while other parameters remains).

It is really true that it is very hard to detect crack using only one frame using traditional image processing method. Pattern recognition using traditional method is usually realized by inputting a video of an object in the similar background. Thenthe frame difference is calculated and the image operatorapplied. And finally the classification of the object out of the background is obtained. But based on our assumption, it is also feasible that after training with a large number of data with very similar backgrounds(coating) but different foreground (crack), the deep learning algorithm can end-to-end learn and find the foreground region using single frame.

Therefore, the accuracy of the algorithm depends largely on the quality of the training dataset, such as the amount of data, the proportion of positive and negative samples, just as other deep learning methods. It should be noted that, if other type of samples are used for evaluation, such as coating corrosion caused by metal particle impurities, coating surface temperature rise caused by gas film blockage, porosity decrease, and coupling thermal effect between multiple cracks, etc., the detection accuracy will be greatly reduced. This is what we want to improve in the future, which is to expand the existing datasets, enhance their generalization, and control the detection error within an acceptable range.”

the authors haven’t addressed the problem of thermal excitation. Or, maybe, I have missed the information as it was hidden behind another data presented. However, since the same thing may have happened to a regular reader, I should ask the question again:

What type of a thermal excitation was applied in the experiment (simulated experiment). In Figs 3 and 5 we observe temperature distributions. What are these temperature rises caused by and are they stationary or transient?

By the way: legend descriptions are still difficult to read.

Author Response

Please see the attachment, thank you.

Author Response File: Author Response.docx

Reviewer 6 Report

- all the abbreviations should be defined at their first use (also in the abstract and in keywords); revisions are needed in several sections of the document

- in figure 1, the box detail picture is not clear enough. Please consider to increase its size of change colours or add labels in support of the reader

Author Response

Please see the attachment,thank you.

Author Response File: Author Response.docx