Depth Evaluation of Tiny Defects on or near Surface Based on Convolutional Neural Network

Round 1

Reviewer 1 Report

The article titled "Depth Evaluation of Tiny Defects on or Near Surface Based on Convolutional Neural Network" presents an interesting approach to detect and evaluate surface micro-defects, especially tiny ones, using Convolutional Neural Networks (CNNs) and laser ultrasonics.

 

1. **Introduction**:

   The introduction effectively highlights the importance of detecting and evaluating tiny defects in critical components. It sets the stage for the research by discussing the challenges in detecting small defects and the potential of CNNs in addressing these challenges.

 

2. **Materials and Methods**:

   - The description of the materials and methods used in the study is comprehensive and provides sufficient detail for replication.

   - The use of laser ultrasonics and wavelet transform is a well-founded approach for defect detection, and the explanation is clear.

   - The optimization of hyperparameters and regularization techniques is essential for deep learning models, and the article covers this aspect well.

 

3. **Results and Discussion**:

   - The graphical representation of training and validation loss curves and accuracy curves is informative and helps readers understand the model's convergence.

   - The achieved accuracy of 91.9% in recognizing defects of various depths is impressive, and the mean deviation of ±0.037mm suggests good precision.

 

4. **Conclusions**:

   - The article concludes by summarizing the key findings and the potential of the proposed method in assessing small defects.

   - It emphasizes the significance of this approach in various industries, which adds practical value to the research.

 

5. **Overall Assessment**:

   - The article is well-structured and provides a clear narrative of the research process.

   - The use of CNNs for defect detection is a relevant and promising application of deep learning.

   - The incorporation of laser ultrasonics and wavelet transform enhances the credibility of the proposed method.

   - The article could benefit from discussing limitations and future research directions, as well as addressing potential challenges in real-world implementation.

 

In summary, the article presents a valuable contribution to the field of defect detection using deep learning and laser ultrasonics. It provides a detailed methodology and achieves impressive accuracy in evaluating small defects. However, some additional discussions and considerations for real-world applications would further strengthen the article.

The English quality in the article is generally good. Here are some key points regarding the quality of English in the article:

 

1. **Clarity**: The article is written in a clear and understandable manner. The sentences and paragraphs are well-structured, making it easy for readers to follow the research process and findings.

 

2. **Technical Terminology**: The article contains technical terminology related to ultrasonics, deep learning, and defect detection. While this is expected in a research paper, the terminology is used appropriately and is explained when necessary, ensuring that both experts and non-experts can grasp the content.

 

3. **Grammar and Syntax**: The grammar and syntax are generally correct. There are very few noticeable grammatical errors, and sentence structures are well-formed.

 

4. **Coherence**: The article maintains a logical flow, connecting ideas and sections smoothly. The use of transitional phrases helps readers navigate through the content.

 

5. **Conciseness**: The article effectively conveys its message without unnecessary verbosity. It presents the research findings and methodology concisely.

 

6. **Academic Style**: The article follows an academic writing style appropriate for a research paper. It uses formal language and adheres to citation and referencing standards.

 

In summary, the English quality in the article is of a high standard, making it accessible to a wide audience, including researchers and professionals in the field.

Author Response

Thank you for your active assessment of our work and valuable suggestions. About how to improve our research to adapt to the needs of different detection scenarios or to get higher detection accuracy is also what we have been trying to do, so additional notes about future research directions have been added in the manuscript (Line 381-387)

Reviewer 2 Report

Dear Authors,

the manuscript is interesting and generally well written.

Here are some comments:

1. Line 38 - please provide examples of surface microdefects.

2. Lines 52-53 - please provide 'significant advancements" which you mentioned.

3. What kind of software was used for neural network preparation?

4. Please clearly indicate the novelty of your research.

Author Response

Thank you for your active assessment of our work and valuable suggestions which greatly help us to improve our manuscript. Point-by-point response to the  comments are listed below:

Q1: Line 38 - please provide examples of surface microdefects

A: Examples of surface microdefects such as surface, near surface cracks or incomplete fusion have been provided (line 38-39)

Q2: Lines 52-53, please provide 'significant advancements" which you mentioned.

A: According to the reviewer’s comment, advancements of the integration of artificial neural networks with ultrasonic detection has been add in the manuscript (Line54-57)

Q3: What kind of software was used for neural network preparation?

A: The neural network preparation was conducted using the TensorFlow.

We have added relevant description in the manuscript (Line 302-303).

Q4: Please clearly indicate the novelty of your research

A: Our research proposes a method for the detection and depth assessment of tiny defects in or near surfaces by combining laser ultrasonics with convolutional neural networks. The novelty of the work can be summarized as follow:

Integration of Wavelet Methods: We address the challenges posed by ultrasonic testing in additive manufacturing by integrating wavelet methods for feature extraction. This enables us to capture key features of signals, particularly crucial in the presence of complex anisotropy and coarse grain noise in additive manufacturing parts.

Optimized Convolutional Neural Network (CNN) Structure: Our work introduces an optimized CNN structure, resulting in improved training efficiency. This enhancement is achieved by refining epochs, ultimately reducing the number of iterations required for convergence. This design choice contributes to the overall advancement of defect detection methodologies.

In summary, by optimizing the dataset (extracting signal features) and optimizing the structure of convolutional neural networks, we have improved training efficiency and achieved competitive accuracy in similar regression models.

Reviewer 3 Report

The topic of the paper is interesting, but the following aspects should be addressed.

1. The size of the detection system seems to be rather inappopriate for its adaptation to various use cases. Is it possible to minify the hardware assembly, without compromising on the detectio performance?

2. It is not clear how the reported approach differs from similar approaches. Therefore, the authors should compare it against 4-5 of the most relevant existing solutions, as they are reported in the referenced literature. This analysis should determine the advantages and drawbacks of the reported solution.

3. The English language should be improved through, at least, one round of proofreading.

The English language should be improved through, at least, one round of proofreading.

Author Response

We appreciate the time and effort you've dedicated to evaluating our work, and point out the omissions in our manuscript. We have revised the manuscript based on your comments.

Q1: The size of the detection system seems to be rather inappropriate for its adaptation to various use cases. Is it possible to minify the hardware assembly, without compromising on the detection performance?

A: The testing space limitation remains a primary hurdle in ultrasound testing for both in-service and in-situ testing requirements. Optimizing the volume of the laser excitation and receiver is a challenging feat. Hence, we aim to design a scanning mechanism that aligns with varied detection scenarios in the future, catering to diverse detection needs. We agree with the reviewer’s comment, minifying the hardware assembly without compromising on the detection performance is one of the aims of our further research and we have added relevant description in “Future research direction”. (Line 382-385).

Q2: It is not clear how the reported approach differs from similar approaches. Therefore, the authors should compare it against 4-5 of the most relevant existing solutions, as they are reported in the referenced literature. This analysis should determine the advantages and drawbacks of the reported solution

 

A: We acknowledge the concern raised regarding the clarity of how our reported approach differs from existing solutions and understand the importance of a comprehensive comparative analysis to highlight the uniqueness of our method. In response to your feedback, we conduct a thorough comparison with some most relevant existing solutions, and our goal is to present a well-rounded discussion that not only showcases the strengths of our approach but also candidly addresses any limitations. Here are the comparisons:

Reference 15 presents an artificial neural network-based approach for identifying defects. The method gathers energy information of various frequency components via wavelet packet decomposition and then feeds it to the ANN neural network to recognize the damage. This research shares a similar research background and methodology to this article, including the use of laser ultrasound to collect data. Composite materials present complex anisotropy, while the coarse grain noise of additive manufacturing parts creates significant challenges for ultrasonic testing. Thus, wavelet methods are utilized to extract vital signal features, followed by machine learning. The proposed model in this literature achieves higher training accuracy (99.6%), but it is limited to distinguishing the presence or absence of defects. However, this study puts forward a regression model which more accurately measures the depth of defects through defect determination. Additionally, the design of the CNN structure leads to a notable decrease in the number of iterations needed for convergence.

In Reference 16, a backpropagation neural network was developed and trained to analyze and predict the sample porosity and the average prediction result is 88.02%, So for quantitative regression models, the depth measurement accuracy we obtain is competitive.

Reference 23 specifically elaborated on the application of CNN convolutional neural networks in detecting defects in additive manufacturing. The parameter design of convolutional and pooling layers, alongside the size of convolutional kernels, have a significant impact on the effectiveness of model training. By analyzing the characteristics of time-frequency plots and the dataset size, we have successfully designed targeted CNN models and achieved competitive training accuracy. However, the accuracy of our model may be enhanced (97%) through physics-based models, transfer learning and other special optimizations, which should be further researched in the future.

Thus, our model differs from other similar domain models due to the dataset preprocessing through continuous wavelet transform and the deliberate design of the convolutional neural network’s convolutional layer, pooling layer, and full-connected layer. This results in a noticeable improvement in the convergence rate the competitive regression accuracy among similar regression models. In order to highlight the above features, we have modified the manuscript. (Line 347-350).

Q3: The English language should be improved through, at least, one round of proofreading

A: We have gone through the language carefully and some descriptions have been revised. (in revised version in the manuscript, for example, Line 46-47)

Round 2

Reviewer 3 Report

Considering its current revised version, the paper can be considered for further processing.

Another round of English language proofreading is advised prior to publication.