Detection and Classification of Bearing Surface Defects Based on Machine Vision

Round 1

Reviewer 1 Report

This paper provided an interesting solution to classify bearing surface defects using machine vision. Several issues need to be addressed and revised before further consideration.

1. Fault diagnosis problems have been successfully addressed by analyzing vibration signals (e.g., Robust Deep Learning-Based Diagnosis of Mixed Faults in Rotating Machinery; Unsupervised Cross-domain Fault Diagnosis Using Feature Representation Alignment Networks for Rotating Machinery). Please compare and discuss the strength of using vision-based technologies instead of analyzing vibration signals.
2. Discussion of state-of-the-art methods in vision-based fault diagnosis are missing, e.g, Deep Learning for Infrared Thermal Image Based Machine Health Monitoring.
3. It would be best if the authors could provide a link to the dataset.
4. Please explain the details of ROI region extraction.
5. Please specify are the images of bearing in Figure 2 collected when the machine is not operating? If so, can the proposed technique be applied to on-line fault diagnosis?
6. Please provide the detailed network architecture in Figure 6, including the inputs and outputs. And explain the type of neural nets used in the proposed architecture. Is it CNN or simply fully connected layers? And the reason for selecting this type of neural nets. Please also indicate how many layers are in the Lc-MNN.
7. Please explain why the Cuboid structure that expands the images to their original size is effective. If the theoretical explanation is not applicable, an ablation study might be needed to empirically validate the Cuboid structure.

 

Author Response

Reviewer 1 Comments

This paper provided an interesting solution to classify bearing surface defects using machine vision. Several issues need to be addressed and revised before further consideration.

1. Fault diagnosis problems have been successfully addressed by analyzing vibration signals (e.g., Robust Deep Learning-Based Diagnosis of Mixed Faults in Rotating Machinery; Unsupervised Cross-domain Fault Diagnosis Using Feature Representation Alignment Networks for Rotating Machinery). Please compare and discuss the strength of using vision-based technologies instead of analyzing vibration signals.

Authors’ Response:

Thank you very much for providing such valuable comments. I carefully read some references and found that many scholars have solved the problems of fault diagnosis by analyzing the vibration signal of the detected object, and achieved good results. With the development of machine vision technology, more and more scholars have done a lot of in-depth and effective work on the surface defect detection of objects such as bearings based on machine vision, such as references [1-2]. Machine vision has the advantages of fast detection speed, a high degree of automation, non-destructive testing, and so on. It has natural advantages in the detection of bearing surface defects. Based on the previous works, this paper focuses on the research of bearing surface defect detection and recognition algorithm based on machine vision. The reviewer’s suggestions will be listed in the future works.

 

2. Discussion of state-of-the-art methods in vision-based fault diagnosis is missing, e.g, Deep Learning for Infrared Thermal Image-Based Machine Health Monitoring.

Authors’ Response:  

Thank you for the reviewer’s suggestion. We have paid attention to some new research methods and achievements in vision-based fault diagnosis, such as references [3-10] cited in the article. In this paper, we proposed a bearing surface defect detection and classification method using machine vision technology. We hope to build an online system, in the choice of imaging scheme,  our focus is to do some useful work in defect detection and classification methods, to achieve rapid detection and classification of bearing defects.

3. It would be best if the authors could provide a link to the dataset.

Authors’ Response:  

I am very willing to share my research data with readers, but all my algorithms are based on the original image data collected by our laboratory, the current data set has about 6.25 G data. I try to upload the data according to the guidelines of the submission website, but the data set is too large, many attempts have failed, and I am very sorry for this. If it's necessary, I'll try to think of other ways. The data used to support the findings of this study are available from the corresponding author upon request.

4. Please explain the details of ROI region extraction.

Authors’ Response:

The bearing is placed flat on the turntable, the linear array camera is parallel to the axis of the bearing, the turntable rotates at a constant speed, and the linear array camera scans the outer ring of the bearing synchronously. The bearing image is divided into three parts: bottom turntable image, middle bearing image, and upper black background. The above three parts are distributed along the Y direction, and the X direction of each part is consistent. The gray level of the black background in the upper part is low, and the gray level of the turntable image in the bottom part is high, but their features are obvious and stable, so it is easy to detect them by using the feature region localization algorithm. ROI of the middle bearing image is extracted by the gray projection algorithm. We added the new details of ROI region extraction are in lines 246 to 254.

5. Please specify are the images of bearing in Figure 2 collected when the machine is not operating? If so, can the proposed technique be applied to on-line fault diagnosis?

Authors’ Response:

The bearing image in Figure 2 can be collected only when the machine is working, and the system can detect it online.

6. Please provide the detailed network architecture in Figure 6, including the inputs and outputs. And explain the type of neural nets used in the proposed architecture. Is it CNN or simply fully connected layers? And the reason for selecting this type of neural nets. Please also indicate how many layers are in the Lc-MNN.

Authors’ Response:

The detailed network architecture in the original figure 6 has been updated on lines 301-303, 305-306, 325, and 369, which is still named Figure 6, which includes the inputs and outputs. We used the type of neural nets used in the proposed architecture is based on BP algorithm, through feedforward neural network, through the training of error backpropagation, completes the nonlinear mapping from the input signal to output mode, obtains the features with high separability through the three-layer screening of features, and eliminates the features with less effect on the classification to achieve dimension reduction, and it is simply fully connected layers. LC MNN in this paper mainly has three layers.

7. Please explain why the Cuboid structure that expands the images to their original size is effective. If the theoretical explanation is not applicable, an ablation study might be needed to empirically validate the Cuboid structure.

Authors’ Response:  

Thank you for the reviewer’s comments on our work. I'm very glad to explain my ideas to you. If it's wrong or not clear, please give me your advice again. When we observe images, if the size of objects is small or the contrast is low, we usually study them with higher resolution; if the size of objects is large or the contrast is high, only rough observation is enough. If there are both smaller objects and larger objects, or objects with lower contrast and higher contrast exist at the same time, it will have advantages to study them with different resolutions. Because the edge of the bearing image has both a fast-changing part and gentle changing part, to meet this demand, the original image is transformed by a three-layer wavelet transform based on the result of 2D wavelet transform, and a series of high-frequency and low-frequency images with pyramid structure is obtained. These images are expanded to the size of the original image by the nearest neighbor interpolation method, and the expansion of the cuboid structure is obtained In addition to the original image, each pixel position is represented by 11-dimensional features, thus the feature vector of each pixel is obtained.

Reviewer 2 Report

The reviewer comments of the paper «Detection and Classification of Bearing Surface Defects Based on Machine Vision»

- Reviewer

The authors presented an article «Detection and Classification of Bearing Surface Defects Based on Machine Vision». In general, the article is interesting and written at a good scientific level. However, there are several points in the article that require further explanation.

Comment 1:

The introduction is well written. However, it will be helpful to add a couple of articles on image processing in section 2. RELATED WORK: doi: 10.1016/j.ymssp.2017.11.022 ; doi: 10.1016/j.ymssp.2016.11.026 .

Comment 2:

3.MATERIALS & METHODS

The quality and resolutions of Figure 2, 3, 4, 7, 8, 9, 10, 11, 12, 13, 14 needs to be improved.

Are the formulas in the article original? If no relevant citations are needed? Explain

Comment 3:

4. EXPERIMENT

Give the chemical composition and physical and mechanical parameters of the bearing materials under study. What characteristics, sizes are used. It is good to give the designations of the bearings and photos.

Provide all measuring devices (manufacturer, city, country).

Comment 4:

It will be useful to add a section of Nomenclature in which to sign all the physical quantities and abbreviations encountered in the article. There are many physical quantities in the text and such a section will help to find the description of the necessary element.

For example,

Lc-MNN              : Local multi-neural network

etc.

 

The topic of the article is interesting. However, authors should carefully study all comments and make improvements to the article step by step. Only after changes can an article be considered for publication in the "Applied Sciences".

Author Response

Reviewer 2 Comments

The authors presented an article «Detection and Classification of Bearing Surface Defects Based on Machine Vision». In general, the article is interesting and written at a good scientific level. However, there are several points in the article that require further explanation.

Comment 1:

The introduction is well written. However, it will be helpful to add a couple of articles on image processing in section 2. RELATED WORK: DOI: 10.1016/j.ymssp.2017.11.022 ; DOI: 10.1016/j.ymssp.2016.11.026 .

Authors’ Response:

Thank you very much for your valuable reference. We have read the above materials carefully and found that neural network has a good application effect in defect image feature detection. We cite these two works to our paper on lines 82-86, 89-93, and put them in our paper as our reference [11] in lines 614-615, [13] in lines 619-620.

Comment 2:

3.MATERIALS & METHODS

The quality and resolutions of Figures 2, 3, 4, 7, 8, 9, 10, 11, 12, 13, 14 need to be improved.

Authors’ Response:

Thank you for your reminder. The high-quality and resolution pictures corresponding to figures 2, 3, 4, 7, 8, 9, 10, 11, 12, 13, and 14 have been packaged and uploaded to the submission system in the form of attachments. Please refer to the attachments.

Are the formulas in the article original? If no relevant citations are needed? Explain

Authors’ Response:

Thank you for your reminder. The formulas in this article are all from references, and formulas (1)-(7) are from reference [46], formulas (8)-(13) are from reference [24] and reference had added in line 402, formulas (14)-(17) are from reference [36] and reference had added in line 415, formula 18 is from reference [32] and reference had added in line 481.

Comment 3:

1. EXPERIMENT

Give the chemical composition and physical and mechanical parameters of the bearing materials under study. What characteristics, sizes are used. It is good to give the designations of the bearings and photos.

Provide all measuring devices (manufacturer, city, country).

Authors’ Response:

The test object of this paper is the outer ring of 6204 deep groove ball bearing (6204 bearings, outer diameter 47 mm, inner diameter 20mm, height 14mm, weight 0.11kg, Cr: 15.8KN, Cor:7.88KN, DW: 9.525mm, Z: 7). The hardware of the detection system mainly includes DALSA linear array camera, linear light source, and self-made loading and unloading control mechanism, and the software of the experiment mainly includes MATLAB, C + +, and so on. The corresponding content has been updated in lines 428-435 of the article.

Comment 4:

It will be useful to add a section of Nomenclature in which to sign all the physical quantities and abbreviations encountered in the article. There are many physical quantities in the text and such a section will help to find the description of the necessary element.

For example,

Lc-MNN              : Local multi-neural network etc.

Authors’ Response:

Thank you very much for your reminder. All the physical quantities and abbreviations in this paper have been described in the paper in lines 590, and there is a special section named “Abbreviations” to list them.

The topic of the article is interesting. However, authors should carefully study all comments and make improvements to the article step by step. Only after changes can an article be considered for publication in the "Applied Sciences".

Authors’ Response:

Thank you very much for your interest in this article and your constructive and friendly suggestions. We carefully read and understand the expert's suggestions and opinions, and seriously revised the original paper, and added relevant supporting materials. We sincerely hope that our revision can accurately express our understanding of the expert's suggestions and opinions. If there are still unclear expressions or other problems, please give us your advice. Thank you again for your suggestions and opinions, and wish you a happy life.

 

Round 2

Reviewer 1 Report

The authors have addressed my previous comments, and the quality of the manuscript has been improved significantly.

Reviewer 2 Report

The authors have improved the article according to the comments. The article can now be published.