Automatic Defect Recognition and Localization for Aeroengine Turbine Blades Based on Deep Learning
Round 1
Reviewer 1 Report
The paper proposes a method for detection of defect on turbine blades that is based on the YOLO family of methods. It is an interesting work that shows promising results. However the following points should be addressed before publication,
1. The reference to AlexNet in page 3, line 97 seems out of place. Better to revise this part, saying that the success of CNNs in classification as in AlexNet has led to their adoption in object detection, or something along these lines.
2. What does the manual feature extraction box in Figure 1 mean? there is no manual extraction in CNNs.
3. I don't understand the added valued of the extended analysis of Section 2 on Faster-RCNN and YOLO. This part can be significantly shortened, a simple reference to the original papers is enough.
4. The following papers should be cited as relevant work, since they describe deep learning methodologies for defect detection using penetrative, non-destructive modalities (x-ray and ultrasound),
a. "ETHSeg: An Amodel Instance Segmentation Network and a Real-world Dataset for X-Ray Waste Inspection" by Qiu et al.
b. "Deep multi-sensorial data analysis for production monitoring in hard metal industry" by Kotsiopoulos et al.
c. "Comparison of deep learning-based image segmentation methods for the detection of voids in X-ray images of microelectronic components" by Schiele et al.
5. Please provide short description on the types of defects (e.g. how are they caused, whether they are internal in the structure of the blade etc.).
6. It would be interesting to know, whether the authors plan to release the studied dataset upon publication, it would help to reproduce their research results.
Author Response
The present authors would like to thank the reviewers and editors for their time and efforts to examine the manuscript and to write very thoughtful comments. We tried our best in answering these questions and, accordingly, in improving the manuscript so that it may be acceptable to Aerospace.
Raised by reviewer 1
- Comment: The reference to AlexNet in page 3, line 97 seems out of place. Better to revise this part, saying that the success of CNNs in classification as in AlexNet has led to their adoption in object detection, or something along these lines.
Response: Many thanks for this comment. We have made the following modifications to this problem (marked by red color).
|
Original manuscript |
Revised manuscript (in page 3, line 97) |
|
2. The object detection algorithm on deep learning The object detection algorithm is …. The accuracy and its real-time performance are relatively low. As shown in the red dashed box in Fig.1. Image classification based on deep learning (convolutional neural network) has received increasing attention with AlexNet proposed in 2012. Thus, object detection based on deep learning subsequently surged. Because of learning features by deep neural networks, .... |
2. The object detection algorithm on deep learning The object detection algorithm is …. The accuracy and its real-time performance are relatively low. As shown in Fig. 1, the red dashed box represents the traditional image processing, and the yellow filled box bellow represents the image processing by CNNs. The success of CNNs in classification as in AlexNet has led to their adoption in object detection. Because of learning features by deep neural networks, .... |
- What does the manual feature extraction box in Figure 1 mean? there is no manual extraction in CNNs.
Response: We are very sorry that we did not give clear explanation for Figure 1. The manual feature extraction box in Figure 1 means some conventional feature extraction methods such as Harris、SIFT、SURF、LBF、HOG et. al. The changes are listed as follows (marked by red color).
|
Original manuscript |
Revised manuscript (in page 3, line 96) |
|
2. The object detection algorithm on deep learning The object detection algorithm is …. The accuracy and its real-time performance are relatively low. As shown in the red dashed box in Fig.1…. |
2. The object detection algorithm on deep learning The object detection algorithm is …. The accuracy and its real-time performance are relatively low. As shown in Fig.1, the red dashed box represents the traditional image processing, and the yellow filled box bellow represents the image processing base on CNNs…. |
- Comment: I don't understand the added valued of the extended analysis of Section 2 on Faster-RCNN and YOLO. This part can be significantly shortened, a simple reference to the original papers is enough.
Response: Section 2 was significantly shortened in revised version.
- Comment: The following papers should be cited as relevant work, since they describe deep learning methodologies for defect detection using penetrative, non-destructive modalities (x-ray and ultrasound),
- "ETHSeg: An Amodel Instance Segmentation Network and a Real-world Dataset for X-Ray Waste Inspection" by Qiu et al.
- "Deep multi-sensorial data analysis for production monitoring in hard metal industry" by Kotsiopoulos et al.
- "Comparison of deep learning-based image segmentation methods for the detection of voids in X-ray images of microelectronic components" by Schiele et al.
Response: Thanks for this comment. We cited these three papers in Section 1 of the revised version. The changes are listed as follows (marked by red color).
|
Original manuscript |
Revised manuscript (line 39, in page 2) |
|
Many scholars devote themselves to the automation and intellectualization of defect detection. Most of these methods are based on computer vision technology for defect detection [8]. In recent years, deep learning has performed well in defect detection [9-11] and segmentation [12, 13] in many industrial fields…. |
Many scholars devote themselves to the automation and intellectualization of defect detection. Most of these methods are based on computer vision technology for defect detection [8]. In recent years, deep learning has performed well in defect detection [9-12] and segmentation [13-16] in many industrial fields.… |
|
References …. 12. Kotsiopoulos, T.; Leontaris, L.; Dimitriou, N.; Ioannidis, D.; Oliveira, F.; Sacramento, J.; Amanatiadis, S.; Karagiannis, G.; Votis, K.; Tzovaras, D.; et al. Deep multi-sensorial data analysis for production monitoring in hard metal industry. The International Journal of Advanced Manufacturing Technology 2021, 115, 823–836. https://doi.org/10.1007/s00170-020-061. …. 15. Qiu, L.; Xiong, Z.; Wang, X.; Liu, K.; Li, Y.; Chen, G.; Han, X.; Cui, S. ETHSeg: An Amodel Instance Segmentation Network and a Real-world Dataset for X-Ray Waste Inspection. InProceedings of the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2283–2. https://doi.org/10.1109/CVPR52688.2022.00232. 16. Schiele, T.; Jansche, A.; Bernthaler, T.; Kaiser, A.; Pfister, D.; Späth-Stockmeier, S.; Hollerith, C. Comparison of deep learning-based image segmentation methods for the detection of voids in X-ray images of microelectronic components. In Proceedings of the 2021 IEEE 17th International Conference on Automation Science and Engineering (CASE). IEEE, 2021, pp. 1. https://doi.org/10.1109/CASE49439.2021.9551671. ….
|
- Comment: Please provide short description on the types of defects (e.g. how are they caused, whether they are internal in the structure of the blade etc.).
Response: Thanks for this comment. We added a short description on the types of defects in Section 3.1.1. The changes are listed as follows (marked by red color).
|
Original manuscript |
Revised manuscript (line 170, in page 6) |
|
…. The 6 kinds of defects were slag inclusion, remainder, broken core, gas cavity, crack and cold shut. Fig. 8 presents these different kinds of defects.… |
…. The 6 kinds of defects were slag inclusion, remainder, broken core, gas cavity, crack and cold shut, which are all internal defects. Slag inclusion, remainder and broken core are all inclusion defects. Looks like there are slag inside of metal castings. Slag inclusion is low-density inclusion, and remainder and broken core are high-density inclusion. Gas cavity is a kind of cavities defect. Gases entrapped by solidifying metal on the inside of the casting, which results in a rounded or oval blowhole as a cavity. Crack of castings can divided into hot crack and cold crack, which are caused by a variety of factors [41]. Some cracks are very obvious and can easily to be seen with the naked eye. Other cracks are very difficult to see without magnification in radiographic testing. Cold shut also called cold lap. It is a crack with round edges. Cold shut is because of low melting temperature or poor gating system. Fig. 7 presents these different kinds of defects.… |
|
References …. 41. Rajkolhe, R.; Khan, J. Defects, causes and their remedies in casting process: A review. International Journal of Research in Advent Technology 2014, 2, 375–383. https://silo.tips/download/defects-causes-and-their-remedies-in-casting-process-a-review#. …. |
- Comment: It would be interesting to know, whether the authors plan to release the studied dataset upon publication, it would help to reproduce their research results.
Response: The studied datasets that support the findings of this study are available from the corresponding author, Hong Xiao, upon reasonable request.
Hong Xiao Email：[email protected]
Author Response File: 
Author Response.docx
Reviewer 2 Report
The paper proposed using deep learning to detect and localize the defects in turbine blades from X-ray images. The work is well-organized and easy to read. Overall, this is good work with enough merits. However, there are several parts that can be improved. Detailed comments are listed below.
1. There are some grammar errors in the manuscripts, which should be corrected carefully.
2. In section 3.1.1, how are these defect X-ray images categorized and labeled? Using certified flawed specimens or labeled by certified and experienced NDT engineers? How to ensure the quality of the original data? This should be included in the corresponding section.
3. In section 3.1.1, the author did not use non-defect images in the training and testing dataset, which is not practical for real engineering applications. The authors could consider how to add the non-defect images in the training and testing datasets
4. Suggest adding a learning curve in section 4.1
5. The numbers of cavity, cold shut, crack are too small in both training and testing datasets, which will make the model performance not convincing (Table 6-8).
Author Response
The present authors would like to thank the reviewers and editors for their time and efforts to examine the manuscript and to write very thoughtful comments. We tried our best in answering these questions and, accordingly, in improving the manuscript so that it may be acceptable to Aerospace.
- Comment: There are some grammar errors in the manuscripts, which should be corrected carefully.
Response: Thanks for this comment. We made detailed language check for spelling, grammar, sentence structures. The changes are marked by red color in revised version.
- Comment: In section 3.1.1, how are these defect X-ray images categorized and labeled? Using certified flawed specimens or labeled by certified and experienced NDT engineers? How to ensure the quality of the original data? This should be included in the corresponding section.
Response: We are sorry for this. In our studied dataset, all defect X-ray images were categorized and labeled by certified and experienced NDT engineers. To ensure the quality of the original data, the original films were scanned with an industrial film digital scanner using optimal resolution scanning parameters. Even so, there are still about 12% of pseudo defects in X-ray images caused by the film itself and the film digitization process. These pseudo defect X-ray images were considered as non-defect. The changes are listed as follows (marked by red color).
|
Original manuscript (line 178, in page 6) |
Revised manuscript |
|
…. Fig. 7 presents these different kinds of defects. For each defective turbine blade picture, label data, including the defect category and the vertex coordinates of defect rectangular area (xmin, ymin, xmax, ymax), were determined by labeling software…. |
…. Fig. 7 presents these different kinds of defects. These defect X-ray images were categorized and labeled by certified and experienced NDT engineers. We found that approximately 12% of the pseudo defects in X-ray images were caused by the film itself and the film digitization process, and these pseudo defects were considered as non-defect. For each defective turbine blade picture, label data, including the defect category and the vertex coordinates of defect rectangular area (xmin, ymin, xmax, ymax), were determined by labeling software…. |
- Comment: In section 3.1.1, the author did not use non-defect images in the training and testing dataset, which is not practical for real engineering applications. The authors could consider how to add the non-defect images in the training and testing datasets?
Response: Thanks for this comment. We are sorry that we have not found a way to use non-defect images in model training at present. Non-defect images are a kind of unlabeled data, which may be applied to unsupervised deep learning. The future work will focus on the research of defect detection algorithm based on unsupervised deep learning for real engineering applications. The changes are listed as follows (marked by red color).
|
Original manuscript |
Revised manuscript |
|
Conclusions and future work …. The current defect detection system still exhibits missed and incorrect detection, especially for detecting defects with small sizes and fuzzy outlines. Further work will focus on small target detection and weak feature information extraction technology. In addition, generative adversarial networks (GANs) [44, 45] will be studied for further data augmentation of defect samples. |
Conclusions and future work …. The current defect detection system still exhibits missed and incorrect detection, especially for detecting defects with small sizes and fuzzy outlines. In real engineering applications, most X-ray images are non-defect. These non-defect big data were not well utilized in our model training and testing. Further work will focus on small target detection and weak feature information extraction technology, and the research of defect detection algorithm based on unsupervised deep learning for real engineering applications. In addition, generative adversarial networks (GANs) [44, 45] will be studied for further data augmentation of defect samples. |
- Comment: Suggest adding a learning curve in section 4.1.
Response: Thanks for this suggestion. We added a learning curve of model training in section 4.1. The changes are listed as follows (marked by red color).
|
Original manuscript |
Revised manuscript |
|
4.1. Model training and testing We selected the …. After that, all layers were unfrozen to train in detail for 50 epochs with a batch size of 8, and the initial training rate was set to 0.0001.
|
4.1. Model training and testing We selected the …. After that, all layers were unfrozen to train in detail for 50 epochs with a batch size of 8, and the initial training rate was set to 0.0001. The learning curve of DBFF-YOLOv4 model is shown in Fig. 14. Figure 14. The learning curve of DBFF-YOLOv4 model.
|
- Comment: The numbers of cavity, cold shut, crack are too small in both training and testing datasets, which will make the model performance not convincing (Table 6-8).
Response: We are sorry that the numbers of defect as cavity, cold shut and crack are too small for such turbine blades in the production. We try our best to collect these datasets. We promise that the model testing results are true and credible.
Author Response File: Author Response.docx
