Automated Defect Detection and Decision-Support in Gas Turbine Blade Inspection
Open accessThe DOI matches its Crossref record; 82 of the 82 references it cites were found in a registry or an index; 19 of the 96 sources its claims cite were checked for support.
Abstract
Background—In the field of aviation, maintenance and inspections of engines are vitally important in ensuring the safe functionality of fault-free aircrafts. There is value in exploring automated defect detection systems that can assist in this process. Existing effort has mostly been directed at artificial intelligence, specifically neural networks. However, that approach is critically dependent on large datasets, which can be problematic to obtain. For more specialised cases where data are sparse, the image processing techniques have potential, but this is poorly represented in the literature. Aim—This research sought to develop methods (a) to automatically detect defects on the edges of engine blades (nicks, dents and tears) and (b) to support the decision-making of the inspector when providing a recommended maintenance action based on the engine manual. Findings—For a small sample test size of 60 blades, the combined system was able to detect and locate the defects with an accuracy of 83%. It quantified morphological features of defect size and location. False positive and false negative rates were 46% and 17% respectively based on ground truth. Originality—The work shows that image-processing approaches have potential value as a method for detecting defects in small data sets. The work also identifies which viewing perspectives are more favourable for automated detection, namely, those that are perpendicular to the blade surface.
Overview
CoChat Summary
TL;DR
This paper develops a heuristic image-processing system to detect edge defects on gas turbine compressor blades and a rule-based decision-support tool that recommends accept/repair/reject actions from engine-manual limits. On a small V2500 blade sample, the detector achieved 83% recall with 54% precision, and the decision tool reduced false positives by 16% while recommending correct maintenance actions in most cases.
Key findings
- In Experiment 2, across 94 images of 47 blades using perspectives P3 and P7, the detector achieved a true-positive/recall rate of 83%, false-negative rate of 17%, precision of 54%, and F1-score of 59%; the abstract reports the combined system detected and located defects with 83% accuracy, with false-positive and false-negative rates of 46% and 17% for a 60-blade sample.
- The best viewing perspectives for model training were 3, 4, 7, and 8, which are the perspectives most perpendicular to the airfoil; increased false positives correlated with decreased true positives and increased false negatives.
- The decision-support tool reduced false positives by 16% by excluding detections at the blade root/dovetail; defect location was determined with 98.4% accuracy and defect depth estimation with 93.7% accuracy, with mean location errors of 3 px/0.9% in x, 8 px/1.3% in y, and 9 px/1.6% radially, and size errors of 9 px/6.3% for depth and 35 px/22.4% for height.
- Detection time was 186–219 ms per blade, compared with an average human inspection time of 85 s for piece-part inspection and about 3 s for borescope inspection; the decision-support tool recommended the correct maintenance action in most cases.
Methods
- Data: 80 high-pressure compressor blades from V2500 engines, stages 6–12, comprising 52 damaged and 28 non-damaged blades; defect proportions were 42% nicks, 38% dents, and 20% tears. Images were acquired with a standardized setup and eight camera perspectives; 33 blades were photographed from eight perspectives (264 images), and 47 blades from P3 and P7 (94 images).
- Detection software: developed in Python 3.7.6 with OpenCV 4.3.0; processing included grayscale conversion, bilateral filtering, Canny edge detection, Suzuki contour extraction, an angle-difference heuristic with a threshold of π/12 rad, a 10×10 region grid, a ground-truth model from 20 non-defective blades, density comparison with a multiplier of 10, and DBSCAN clustering.
- Decision-support tool: a rule-based lookup table using fictional limits by blade stage and zone; it compared defect location and depth against acceptance, repair, and reject thresholds, with performance evaluated against ground truth from a senior inspector with over 30 years of experience.
Limitations
- Stated: the decision-support tool treats all detected defects as dents and nicks and cannot process tears, which always require rejection; defect classification is future work. The detector had difficulty with long smooth edge defects whose angle change fell below threshold, and although it detected all teared-off corners, computed bounding boxes were significantly smaller than actual defects, causing size and location discrepancies. The model comparison module uses only point densities, so shape is not represented, and defects near grid-region edges may be split and missed. No rare crack defect was present in the sample, data are not publicly available due to commercial sensitivity, and the maintenance limits are fictional.
- Evident: the dataset is small (80 blades; 94-image test set), and the paper reports no direct same-dataset comparison between software and human inspector performance. The decision-support tool only processes positive detections from the detector (true positives and false positives), so false negatives are not handled.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

![Figure 2 Cross section of a V2500 gas turbine engine with highlighted compressor blades in yellow and exposed sample. Image adapted from [12]. Figure 2. Cross section of a V2500 gas turbine engine with highlighted compressor blades in yellow and exposed sample. Image adapted from [12]. resulting in material separation and breakage of the entire blade under centrifugal load [14], which has the potential to cause severe damage to the engine and aircraft.](https://figures.cochat.ai/paper-figures/677b258b296340a1b5594bbbe81d3125/automated-defect-detection-and-decision-support-in-gas-turbine-blade-inspection-figure-2.png)













Verification
Claims the paper makes
14 Valid · 17 Not valid · 24 Not checked · 99 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Hence, the neural network method critically depends on relatively larger training and test datasets compared to image processing methods [52-54].”
What the sources say
- Not checked — no full text of the sourcecites [52] 52. Chu, C.; Hsu, A.L.; Chou, K.H.; Bandettini, P.; Lin, C. Does feature selection improve classification accuracy? Impact of sample size and feature selection on classification using anatomical magnetic resonance images. Neuroimage 2012 , 60 , 59-70. (In English) [CrossRef] [PubMed] source
- Partly supportedcites [53] 53. Cho, J.; Lee, K.; Shin, E.; Choy, G.; Do, S. How much data is needed to train a medical image deep learning system to achieve necessary high accuracy. arXiv 2015 , arXiv:1511.06348v2. sourceExperimental — accuracy not yet measured
The Results and Discussions (p. 6) report that larger training sets improved this CNN’s classification accuracy, supporting the training-data portion. The paper does not establish that neural networks require larger training or test datasets than image-processing methods; it reports using 6,000 test images but makes no comparative claim about test-set requirements. (LLM, conf=0.86)
- Not checked — no full text of the sourcecites [54] 54. Foody, G.M. Sample size determination for image classification accuracy assessment and comparison. Int. J. Remote Sens. 2009 , 30 , 5273-5291. [CrossRef] source
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“There is a lack of recent applications of these techniques to blade inspection. In fact, the field is somewhat weak and has been dominated by the neural networks and deep learning approaches instead [9,19,47,85].”
What the sources say
- Not checked — no full text of the sourcecites [9] 9. Shen, Z.; Wan, X.; Ye, F.; Guan, X.; Liu, S. Deep Learning based Framework for Automatic Damage Detection in Aircraft Engine Borescope Inspection. In Proceedings of the 2019 International Conference on Computing, Networking and Communications (ICNC), Honolulu, HI, USA, 18-21 February 2019; pp. 1005-1010. source
- Could not verifycites [19] 19. Malekzadeh, T.; Abdollahzadeh, M.; Nejati, H.; Cheung, N.-M. Aircraft Fuselage Defect Detection using Deep Neural Networks. arXiv 2017 , arXiv:1712.09213v2. sourceExperimental — accuracy not yet measured
The Introduction (p. 1) discusses a lack of prior automatic image-based aircraft defect detection, but the paper studies fuselage defects, not blade inspection. Searches for blade inspection and classical image-processing applications found no passage assessing whether recent blade-inspection work using those techniques is scarce. (LLM, conf=0.78)
- Not checked — no full text of the sourcecites [47] 47. Bian, X.; Lim, S.N.; Zhou, N. Multiscale fully convolutional network with application to industrial inspection. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Placid, NY, USA, 7-9 March 2016; pp. 1-8. source
- Could not verifycites [85] 85. Svensen, M.; Hardwick, D.; Powrie, H. Deep Neural Networks Analysis of Borescope Images. In Proceedings of the European Conference of the PHM Society, Utrecht, The Netherlands, 3-6 July 2018. sourceExperimental — accuracy not yet measured
The Abstract describes this paper’s use of deep neural networks for borescope images, and the Conclusion and Discussion mentions another neural-network application to turbine-blade images. Neither passage assesses how many recent blade-inspection studies use classical image processing. Searches for classical and traditional image-processing approaches found no such assessment in our copy. (LLM, conf=0.73)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, they require significant amounts of data to accurately perform a classification, especially with increasing number of classes [52-54].”
What the sources say
- Not checked — no full text of the sourcecites [52] 52. Chu, C.; Hsu, A.L.; Chou, K.H.; Bandettini, P.; Lin, C. Does feature selection improve classification accuracy? Impact of sample size and feature selection on classification using anatomical magnetic resonance images. Neuroimage 2012 , 60 , 59-70. (In English) [CrossRef] [PubMed] source
- Partly supportedcites [53] 53. Cho, J.; Lee, K.; Shin, E.; Choy, G.; Do, S. How much data is needed to train a medical image deep learning system to achieve necessary high accuracy. arXiv 2015 , arXiv:1511.06348v2. sourceExperimental — accuracy not yet measured
The Learning Curve section (p. 4) says classification accuracy depends heavily on training-set size, and Results and Discussions (p. 6) estimates that this classifier needs 4,092 images per class to reach its target accuracy. The study uses six fixed anatomical classes; it does not test whether increasing the number of classes increases the amount of data needed. (LLM, conf=0.86)
- Not checked — no full text of the sourcecites [54] 54. Foody, G.M. Sample size determination for image classification accuracy assessment and comparison. Int. J. Remote Sens. 2009 , 30 , 5273-5291. [CrossRef] source
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In contrast, the minimum number of images for traditional image processing approaches is much less, of the order of about 10-100 [82,83].”
What the sources say
- Partly supportedcites [82] 82. Kim, K.-B.; Woo, Y.-W. Defect Detection in Ceramic Images Using Sigma Edge Information and Contour Tracking Method. Int. J. Electr. Comput. Eng. (IJECE) 2016 , 6 , 160. [CrossRef] sourceExperimental — accuracy not yet measured
The Experiment and Result Analyses section reports testing the proposed image-processing method on 12 ceramic images, a number within the claimed range. It does not say that 12—or 10–100—is the minimum number required, or establish a minimum for traditional image-processing approaches generally. (LLM, conf=0.78)
- Not matchedcites [83] 83. [Nishu; Agrawal, S. Glass Defect Detection Techniques Using Digital Image Processing-A Review. 2021. Available online: https://www.researchgate.net/profile/Sunil\_Agrawal6/publication/266356384\_Glass\_Defect\_Detection\_Techniques\_ using\_Digital\_Image\_Processing\_-A\_Review/links/5806f81008ae03256b76ff84.pdf (accessed on 10 January 2021).](https://www.researchgate.net/profile/Sunil_Agrawal6/publication/266356384_Glass_Defect_Detection_Techniques_using_Digital_Image_Processing_-A_Review/links/5806f81008ae03256b76ff84.pdf)
We searched by its title and found no matching paper.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The detection rate is comparable with the performance of neural networks in the reviewed literature [21,35,42,48], which ranged from 64.4% to 85%.”
What the sources say
- Partly supportedcites [21] 21. Dogru, A.; Bouarfa, S.; Arizar, R.; Aydogan, R. Using Convolutional Neural Networks to Automate Aircraft Maintenance Visual Inspection. Aerospace 2020 , 7 , 171. [CrossRef] sourceExperimental — accuracy not yet measured
Appendix A, Table 3 reports recall values for this paper’s neural-network dent-detection experiments, including 64.08% and 69.97%, providing an example broadly within the claimed range. It does not establish that detection rates across the reviewed literature ranged from 64.4% to 85%, or explicitly make the claimed comparison. (LLM, conf=0.78)
- Not matchedcites [35] 35. Tian, W.; Pan, M.; Luo, F.; Chen, D. Borescope Detection of Blade in Aeroengine Based on Image Recognition Technology. In Proceedings of the International Symposium on Test Automation and Instrumentation (ISTAI), Beijing, China, 20-21 November 2008; pp. 1694-1698.
We searched by its title and found no matching paper.
- Not checked — no full text of the sourcecites [42] 42. Shipway, N.J.; Barden, T.J.; Huthwaite, P.; Lowe, M.J.S. Automated defect detection for Fluorescent Penetrant Inspection using Random Forest. NDT E Int. 2018 , 101 . [CrossRef] source
- Partly supportedcites [48] 48. Wang, K. Volume CT Data Inspection and Deep Learning Based Anomaly Detection for Turbine Blade. Ph.D. Thesis, University of Cincinnati, Cincinnati, OH, USA, 2017.Experimental — accuracy not yet measured
The results table (Additional information, pp. 45–49) gives CNN2 an average error rate of 14.8991%, corresponding to approximately 85% accuracy. It does not give a 64.4% detection rate or establish the stated 64.4%–85% range; Section 3.7 (pp. 50–52) reports error rates below 0.5% for other network tests. (LLM, conf=0.79)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
50 more claims
Create a free account to see all 154 claims and to verify the 5 references still open.
- Probably not a substantial claim. 70
- It may state two separate claims as one. 52
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.2.2. Decision Output and Recommended Maintenance Action · p. 18
“The defect size had a computation error of 9 pixels or 6.3% for the depth and 35 pixels or 22.4% for the height.”
§ 4.2.2. Decision Output and Recommended Maintenance Action · p. 19
“the defect height had a much bigger error with a mean discrepancy of 29%.”
Citations
Cites
Distinctive Image Features from Scale-Invariant Keypoints
Lowe, D. G. (2004) · International Journal of Computer Vision · doi:10.1023/b:visi.0000029664.99615.94 · Cited by 55,589 · cited 1×
“The approach used the scale invariant feature transform (SIFT) algorithm [44,45] and principal component analysis to produce a damage representation, which is then compared with input images.”
Method(1996) · Cited by 20,179
Object recognition from local scale-invariant features
Lowe, D. G. (1999) · Proceedings of the Seventh IEEE International Conference on Computer Vision · doi:10.1109/iccv.1999.790410 · Cited by 16,472 · cited 1×
“The approach used the scale invariant feature transform (SIFT) algorithm [44,45] and principal component analysis to produce a damage representation, which is then compared with input images.”
MethodTopological structural analysis of digitized binary images by border following
Suzuki, S., & be, K. (1985) · Computer Vision, Graphics, and Image Processing · doi:10.1016/0734-189x(85)90016-7 · Cited by 2,663 · cited 1×
“The Suzuki algorithm [88] is then applied to the found edges in order to extract contours and order them in a hierarchical structure.”
MethodRevisiting Unreasonable Effectiveness of Data in Deep Learning Era
Sun, C., Shrivastava, A., Singh, S., & Gupta, A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.97 · Cited by 2,484 · cited 1×
“[77] used 300 million images to train a neural network.”
MethodLearning methods for generic object recognition with invariance to pose and lighting
LeCun, Y., Huang, F. J., & Bottou, L. (2004) · Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. · doi:10.1109/cvpr.2004.1315150 · Cited by 1,324
A universal noise removal algorithm with an impulse detector
Garnett, R., Huegerich, T., Chui, C., & He, W. (2005) · IEEE Transactions on Image Processing · doi:10.1109/tip.2005.857261 · Cited by 615 · cited 1×
“Of the image-processing algorithms, positive results have been shown for bilateral filtering and Gaussian blur algorithms [49,50].”
MethodOn Loss Functions for Deep Neural Networks in Classification
Janocha, K., & Czarnecki, W. M. (2017) · Schedae Informaticae · doi:10.4467/20838476si.16.004.6185 · Cited by 591 · cited 1×
“Typically, neural networks are used for classification tasks [51].”
MethodRecent Advances in Active Infrared Thermography for Non-Destructive Testing of Aerospace Components
Ciampa, F., Mahmoodi, P., Pinto, F., & Meo, M. (2018) · Sensors · doi:10.3390/s18020609 · Cited by 568 · cited 1×
“This encompasses mainly crack detection [9,33,37,38,42], as this is the most critical type of defect and the main source of failure.”
BackgroundMachine learning to classify animal species in camera trap images: Applications in ecology
Tabak, M. A., Norouzzadeh, M. S., Wolfson, D. W., Sweeney, S. J., Vercauteren, K. C., Snow, N. P., Halseth, J. M., Di Salvo, P. A., Lewis, J. S., White, M. D., et al. (2019) · Methods in Ecology and Evolution · doi:10.1111/2041-210x.13120 · Cited by 504
Research works printed but not linked: 7
- Rao, 2014, Failure mechanisms in turbine blades of a gas turbine Engine—An overview, Int. J. Eng. Res. Dev., 10, 48
- Dewangan, 2015, Gas turbine blades—A critical review of failure at first and second stages, Int. J. Mech. Eng. Robot. Res., 4, 216
- 29. Lufthansa Technik. Mobile Robot for Fuselage Inspection (MORFI) at MRO Europe. Available online: http://www. lufthansa-leos.com/press-releases-content/-/asset\_publisher/8kbR/content/press-release-morfi-media/10165 (accessed on 2 November 2019).
- 35. Tian, W.; Pan, M.; Luo, F.; Chen, D. Borescope Detection of Blade in Aeroengine Based on Image Recognition Technology. In Proceedings of the International Symposium on Test Automation and Instrumentation (ISTAI), Beijing, China, 20-21 November 2008; pp. 1694-1698.
- 83. [Nishu; Agrawal, S. Glass Defect Detection Techniques Using Digital Image Processing-A Review. 2021. Available online: https://www.researchgate.net/profile/Sunil\_Agrawal6/publication/266356384\_Glass\_Defect\_Detection\_Techniques\_ using\_Digital\_Image\_Processing\_-A\_Review/links/5806f81008ae03256b76ff84.pdf (accessed on 10 January 2021).](https://www.researchgate.net/profile/Sunil_Agrawal6/publication/266356384_Glass_Defect_Detection_Techniques_using_Digital_Image_Processing_-A_Review/links/5806f81008ae03256b76ff84.pdf)
- 92. Cohen, S. Measuring Point Set Similarity with the Hausdorff Distance: Theory and Applications. Ph.D. Thesis, Stanford University, Stanford, CA, USA, 1995.
- 97. Drury, C.G.; Fox, J.G. The imperfect inspector. In Human Reliability in Quality Control ; Halsted Press: Sydney, Australia, 1975; pp. 11-16.
Cited by
Alatorre, D., Robles-Linares, J. A., Russo, M., et al. (2022) · IEEE Robotics & Automation Magazine · doi:10.1109/mra.2022.3223220 · Cited by 85
“For example, the components within a gas turbine can only be accessed through narrow channels, such as borescope ports and a reach of several meters [1], [2].”
BackgroundZhang, J., Cosma, G., & Watkins, J. (2021) · Journal of Imaging · doi:10.3390/jimaging7030046 · Cited by 93
Shang, H., Wu, J., Sun, C., Liu, J., Chen, X., & Yan, R. (2022) · IEEE Transactions on Industrial Informatics · doi:10.1109/tii.2022.3222300 · Cited by 66
Surface Defect Detection of Aeroengine Blades Based on Cross-Layer Semantic Guidance
Song, K., Sun, X., Ma, S., & Yan, Y. (2023) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2023.3276026 · Cited by 71
“Aust et al. [25] proposed an image processing method for the automatic detection of blade damage.”
MethodAust, J., & Pons, D. J. (2022) · Applied Sciences · doi:10.3390/app12042250 · Cited by 69
Self-citation
Metadata
- Authors
- Aust, J., Shankland, S., Pons, D., Mukundan, R., & Mitrovic, A.
- Affiliations
- University of Canterbury (NZ)
- Department of Computer Science and Software Engineering, University of Canterbury, Christchurch 8041, New Zealand
- Corresponding author
- Aust, Jonas
- Venue
- Aerospace
- Publisher
- MDPI AG
- Published
- 2021-01-26
- Volume
- 8
- Issue
- 2
- Pages
- 30
- ISSN
- 2226-4310 (ISSN-L)
- OpenAlex
- W3125286443
- Cited by
- 66
- References
- 82
- Citations by year
- 2026: 13 · 2025: 15 · 2024: 13 · 2023: 5 · 2022: 13 · 2021: 7
- FWCI
- 4.92
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Aerospace (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- University of Canterbury Research Repository (University of Canterbury) (submittedVersion)
- Funders
- Christchurch Engine Centre (2018-01)
- Topics
- Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
- Physical Sciences › Engineering › Mechanical Engineering › Non-Destructive Testing Techniques
- Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
- Keywords
- automated defect detection, image processing, small datasets, decision support, maintenance recommendations, visual perspective, aircraft engine maintenance
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-12, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-12 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.