Purpose: Nicks in the propeller can lead to catastrophic failures if not repaired properly, but proper detection can be difficult. To improve nick inspection tasks, the researchers intend to develop a nick detection model using the object detection algorithm, You Only Look Once.

Design/Methodology/Approach: The researchers select YOLOv8 algorithm. To develop a nick detection model, the researchers collected 4423 nick images in various conditional variable settings. Using preprocessed 1800 nick images, six models are trained with different hyperparameter settings and assessed for performance.

Findings: A well-trained model achieved 0.986 precision, 0.982 recall, and 0.984 F1 score. Additionally, the researchers assessed explainability analysis using Grad-cam to see if this model properly extracts nick features during the training process.

Originality/Value: This research contributes to improving inspectors’ propeller nick inspection maturity using this nick detection model. Therefore, this study can potentially reduce general aviation accidents caused by propeller failure.

1. Introduction

In powered flight, there is arguably not a more critical part in the production of thrust than the propeller. Proper maintenance practices must be followed because the propeller is not only essential for flight, but it is also under a tremendous amount of stress from various sources such as centrifugal force, torque, thrust, and aerodynamic loads [1]. The propeller also operates in dynamic operating conditions. Stresses and operating conditions can lead to defects on the surface of the propeller, such as nicks or cracks, which can lead to the failure of the propeller [1].

Finding and repairing defects while maintaining a smooth, aerodynamic surface are key inspection and maintenance practices. Mechanics and pilots perform visual inspections to ensure the propeller is kept in an airworthy condition. If a defect is found, repairs must be made by a Federal Aviation Administration (FAA)–certified powerplant maintenance technician. Eligibility requirements can be found in 14 CFR Part 65 [2], maintenance practices and scopes in 14 CFR Part 43 [3], and maintenance details in Advisory Circular (AC) 20-37 [1]. The current maintenance practices include both corrective maintenance and preventive maintenance. Corrective maintenance is performed when a defect is detected and restored to an airworthy condition. Preventive maintenance, such as minor preservation operations or the replacement of small standard parts, is generally performed based on certain periodic or time-based schedules. In addition, aircraft operators perform both pre- and postflight inspections around aircraft operations.

Since propeller failure can result in injury, death, or significant equipment failure, it is important for pilots, operators, and technicians to properly identify and repair possible sources of fatigue damage, such as a nick. Machine learning (ML) has been widely adopted in various domains as a decision supporter to overcome decision errors produced by humans. Depending on the ML algorithms, they can generate outputs that provide explanations, predictions, or definitions of certain situations [4]. The principle of ML is to train a computer model on various patterns and characteristics from provided samples. The trained model can recognize objects that look similar to the shape of the sample [5]. In this research, the researchers intend to develop a nick detection model using the object detection algorithm, You Only Look Once (YOLO).

2. Materials and Methods

2.1. YOLO

YOLO is a convolution neural network (CNN)–based ML algorithm; the first generation is developed by [6], and the YOLO outperforms two other well-structured detection algorithms, (a) deformable parts model (DPM) and (b) fast region-based convolutional neural network (R-CNN), in terms of speed [6, 7]. As of January 2023, YOLOv8 was released by Jocher et al. [8]. Jocher et al. [8] establish that YOLOv8 leads in both accuracy and speed compared to its previous YOLO versions. The YOLOv8 also provides five different scale models (nano (n) = smallest, s = small, m = medium, large (l) = large, and x-large (x) = largest). As scale size decreases, training time and inference speed become faster, but prediction accuracy decreases. Conversely, as the scale size increases, training time, inference speed, and prediction accuracy also increase.

While the in-depth architecture of the YOLOv8 is covered by these studies [7, 9, 10], the YOLOv8 architecture consists of four main bodies: backbone, neck, head, and loss, as shown in Figure 1:

-
Backbone: The YOLOv8 selects cross stage partial (CSP) [12] that extracts the main input features. Spatial pyramid pooling feature (SPPF) and channel fusion (C2f) are used to improve computational processes and increase accuracy.
-
Neck: The neck acts as a bridge between the backbone and head to make better prediction results. The neck process refines aggregated data from the backbone data. It reprocesses it by focusing on geometric and conceptual information based on two scales: feature pyramid networks (FPNs) [13] and path aggregation networks (PANs) [14].
-
Head: The head is where the object prediction is processed based on a given new input determined by the extracted features from the backbone and neck. YOLOv8 adopts anchor-free improved approaches that help find the location of new input.
-
Loss: The loss evaluates errors between the predicted location and actual location and can be evaluated into three losses: distribution focal loss (DFL), classification loss (CLS), and bounding box (BOX) losses.

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

Typical YOLOv8 architecture diagram [9, 11].

2.2. ML Implementation Studies

The researchers review ML implementation studies for propellers or blades in the aeronautical domains using borescope inspection tasks to detect defects in gas turbine blades [15–21] and in propellers [22, 23]. Furthermore, the researchers review other YOLO implementation case studies: inspection of steel strip surfaces using YOLOv1 which can detect scar, scratch, inclusion, burr, seam, and iron scale [24]; a YOLOv3 model that can detect aircraft skin defects [25]; Zhang et al. [26] develop turbine engine blade defects’ detection model using YOLOv3; a wind turbine blade defect detection model using YOLOv5 [17, 27, 28].

2.3. Propeller Failure

While there are various defect types on the propeller identified by AC 20-37 [1], this research is focused on the nick, a “sharp, notchlike displacement of metal usually found on leading and trailing edges” of propellers ([1], p. 3). Common propeller failures that begin in the tip are frequently caused by a nick on the propeller surface [1, 29]. Caldwell [30, 31] states that a nick is a tip failure initiator; when the aircraft starts, the propeller stirs objects or material on the ground, and the propeller strikes the object or material and creates a nick. Furthermore, when the propeller fails due to nick, it does not show any structural weaknesses or internal flaws [29, 30]. This means nick damage on the propeller plays a significant role in propeller failure, rather than inherent weaknesses or design flaws. Therefore, finding nicks is a crucial task and should be found before aircraft take-off [29–33].

Once a nick is found, it is also essential to repair the propeller as soon as possible [31]. The general steps to repair a nick are to round off sharp areas using a file. Then, the technician removes tool marks and polishes the propeller with ever increasing grits of emery and crocus cloth before inspecting with a magnifying glass. The detailed propeller repair steps, tools, and consumables are described in the original equipment manufacturers’ (OEM) maintenance procedures, or in the absence of OEM manuals, sometimes technicians may use AC 20-37 as approved instructions.

Improper inspection and maintenance can result in a failure to identify and repair nicks, cracks, and dents that can result in the catastrophic failure of the propeller. When propellers fail, injury and damage to the aircraft are substantial. The inability to identify nicks can be influenced by different factors such as risk sensitivity, individual personality, and human error factors [34, 35], including the lack of experience and knowledge. New technologies, such as ML, can be used to support pilots and technicians to identify nicks that may previously have been overlooked in propellers so they can be properly repaired.

3. Problem Generation

The ability of an inspector to perform a proper inspection can vary based on multiple factors. To improve propeller inspection outcomes, the researchers intend to develop a nick detection model using YOLOv8 algorithm to support inspectors in their propeller inspection tasks. This model can detect nicks that inspectors may miss during inspections.

4. Model Development

To develop a nick detection model, the researchers follow the model development workflow as shown in Figure 2. This workflow includes research design and results. In this model development section, nick images are collected, preprocessed, and the YOLOv8 is selected, and then, six different hyperparameters are selected to train and validate the YOLOv8 models. Next, in the Results section, among the six YOLOv8-M models, the researchers determined a well-performed nick detection model based on the validation results.

4.1. Nick Image Collection

Since obtaining actual nick images is challenging, the researchers create a test coupon from an unairworthy propeller blade used in an educational setting to train the models with images that closely resemble actual nicks. To achieve this, the researchers create six nicks in the blade using a hammer and aluminum chisel; see Figure 3.

In the nick image collection process, the researchers consider three conditional variables: standpoint range, view angles, and lighting across the six nicks. These three variables reflect the inspector’s dynamic perspective. The different angles of the nick images help the models understand a diverse range of perspectives of the nicks during the training process.

-
View angles (a): The view angle refers to the camera’s vertical position relative to the propeller, captured at three angles: 90°, 45°, and 0°, as shown in Figure 4.
-
Standpoint range (b): Standpoint refers to the camera’s location on the horizontal plane relative to the propeller. This ranged from 0° to 180°, as shown in Figure 4.

The shape of a nick can vary depending on the aircraft inspectors’ diverse standpoint and view angles. To account for this variability, the researchers capture images of each nick across the full range of standpoints (0°–180°) and all three view angles (90°, 45°, and 0°). This is accomplished by using the camera’s continuous shooting mode while moving the camera smoothly from left to right. An example of this image collection approach for one nick, incorporating both standpoint and view angles, is shown in Figure 4.

-
Light: The inspectors typically use a flashlight to perform night inspections. When a flashlight is shined at the surface of the propeller, the nick reflects a different shine if a nick is present. Inspectors typically do not use flashlights for daytime inspections. Therefore, two light conditions are applied to the view angles and standpoint range variables: when there is no light, the nick images are collected, and in a dark environment, the flashlight is given at the nick, with the reflected nick images being collected.

Considering all conditional variables on the six nicks, the researchers collected 4423 images in total. The number of nick images under each condition is provided in Table 1.

Table 1. Collected nick images for each conditional variable.

Nick	#1			#2			#3
Standpoint	0°–180°			0°–180°			0°–180°
View angles	90°	45°	0°	90°	45°	0°	90°	45°	0°
Light	99	127	113	158	88	116	106	135	119
No light	112	65	133	112	113	117	126	122	128
Nick	#4			#5			#6
Standpoint	0°–180°			0°–180°			0°–180°
View angles	90°	45°	0°	90°	45°	0°	90°	45°	0°
Light	100	109	113	169	134	184	158	115	110
No light	121	138	92	134	113	144	137	131	132

4.2. Nick Image Collection Preprocessing

To prepare the nick models’ training, the researchers conduct nick image collection preprocessing. This preprocessing is composed of three tasks and is required for training and validation:

-
Random subsampling: Various nick images are collected in conditional variables as shown in Table 1. However, due to the limited computational resources, the researchers use Python random function to randomly subsample 50 images for each nick and conditional variables.
-
Image labeling: Since the YOLO algorithm is supervised learning, it requires manual labeling of nick boundaries in all subsampled images. The authors, who hold FAA-certified Airframe and Powerplant licenses, manually labeled the boundaries on nick images as class “NICK” using V7-Labs [36] software.
-
Image partition: Using the same Python random function, 40 images were selected out of the 50 randomly selected images from the random subsampled images for training the models. The remaining 10 images are selected for validating models. Therefore, 1440 images are allocated for training, and 360 images are allocated for validation.

4.3. Model Selection

The YOLOv8 performs fast inference speed and high accuracy rates compared to previous YOLO versions [8]. Furthermore, the YOLOv8 model supports image augmentation that helps to better fit the model. Among five YOLO scale models, a YOLOv8-M is selected, because that is well balanced between accuracy and prediction speed [8].

4.4. Hyperparameters

Since different hyperparameters can affect the performance of each model, three hyperparameters are used to train the YOLOv8-M model: learning rate, epoch, and batch, and these three hyperparameters are used in different settings to find an optimized model.

The researchers select a grid search [37] for two hyperparameters: batch and learning rate, which the grid search technique involves inputting different values to explore the performance of various parameter combinations. The epoch size is automatically optimized through the early stopping algorithm, which can prevent overfitting [9, 23] and waste computational resources from redundant training.

The researchers use a stochastic gradient descent (SGD) optimizer and set batches to 16 and 32 and set the learning rate to 1 × 10⁻², 1 × 10⁻³, and 1 × 10⁻⁴. The early stopping algorithm sets to patience of 5. If there is no model improvement in the model based on the validation weight results after each epoch within the given patience, the training process stops, and no future epoch is processed [8]. For example, all training images are used to train in the first epoch, and after the validation images are used to validate model performance using metrics discussed in Section 4.5. If there is no model performance improvement in validation results for five consecutive iterations of the epoch, the model retrieves the best-performing state from the last five epochs. If the model shows continuous improvement in validation results, it is trained continuously. The trained models’ hyperparameters are shown in Table 2. Note: The number of epochs for each model is determined based on the validation results after each training epoch. Models (A) and (D) indicate that one epoch is an optimized epoch size; no performance improvements are observed during five consecutive iterations of the epoch. Models (B) and (E) determine 41 and 31 epoch sizes, respectively. Models (C) and (F) determine 27 and 25 epoch sizes, respectively.

Table 2. Trained Model Hyperparameters & Validation Results.

Model	Hyperparameters			Validation results
Model	Batch	Learning rate	Epoch	Precision	Recall	mAP @50	mAP @50–95	F1score
(A)	16	1 × 10⁻²	1	0.887	0.828	0.867	0.386	0.856
(B)	16	1 × 10⁻³	41	0.986	0.982	0.986	0.595	0.984
(C)	16	1 × 10⁻⁴	27	0.958	0.958	0.975	0.526	0.958
(D)	32	1 × 10⁻²	1	0.861	0.793	0.847	0.381	0.826
(E)	32	1 × 10⁻³	31	0.983	0.983	0.983	0.579	0.983
(F)	32	1 × 10⁻⁴	25	0.966	0.953	0.974	0.527	0.959

Note: Bold indicates the final model (B)’s hyperparameters and results.

4.5. Metrics

For this study, the six models are evaluated using four metrics [38, 39]: precision (P), recall (R), F1 score, and accuracy rate. Precision is the confidence or proportion of true positive (TP) in actual images classified as positive over a summation of TP and false positive (FP); see Equation (1). Recall is the sensitivity or proportion of TPs in predictive images classified as true over summation of TP and false negative (FN); see Equation (2). Equation (3) is the F1 score equation to indicate the harmonic average of recall and precision. Equation (4) is the intersection of union (IoU) equation. The IoU variables are denoted as B_gt which is the bounding box ground truth and B_p which is the bounding box prediction; the high ratio interaction between two boxes indicates a high IoU. Equation (5) represents the average precision (AP) of each class [40]. Equation (6) represents the mean AP (mAP) of all classes at [IoU threshold]. Equation (7) represents validation weights, which are the sum of validation metrics used to determine the model performance improvement at each epoch for the early stopping algorithm [8]. This research evaluates the model with two IoU thresholds: 50 and 50–95. Additionally, the researchers use a Grad-cam algorithm, which is a local explanation methodology [41, 42], to visualize how CNN extracts the features from training images. This algorithm is applied to a final selected optimized model to provide an explainability analysis, evaluating whether the model properly extracted nick features during the CNN training process.

()

5. Results

The performance of each model is evaluated using validation images, and the results are shown in Table 2. While Models (A) and (D) perform noncompetitive performance across all metrics compared to Models (B), (C), (E), and (F), Model (B) outperforms four evaluation assessments: P, mAP at 50, mAP at 50–95, and F1 score, and Model (E) outperforms only R.

5.1. Determine an Optimized Model

Based on the validation results, the researchers select Model (B) as the final model. The optimized hyperparameters are 16 batches with 1 learning rate and 41 epochs. This is because Model (B) performs at a higher F1 score, which is harmonically well balanced for minimizing FPs through high precision and capturing more TPs through high recall. Therefore, Model (B) is most suitable for propeller inspection tasks. For example, higher precision indicates fewer FPs, meaning the model is less likely to make incorrect inspections that can lead to redundant maintenance. Although Model (B) shows a 0.001 decrease in precision compared to Model (E), it maintains its ability to correctly identify actual TPs while minimizing FNs, such as actual defects are not overlooked or incorrectly classified as nondefective.

Model (B) training summary is shown in Figure 5. Precision–recall curve indicates 0.986 is AP at 50 of “NICK” class (refer to Figure 6a). Figure 6b shows x and y coordinates of normalized images are normally distributed, and the width and height are exponentially distributed. In addition, a label correlogram is provided to show the statistical distribution with image label variables’ pairwise correlations. The x and y coordinates represent the center point of the bounding box; the width and height represent the length of the bounding box relative to image size [6, 43–45]. The results of the validation sample images are shown in Figure 7.

5.2. Explainability Analysis

The researchers use a Grad-cam algorithm on Model (B) to verify how the YOLOv8-M model extracts nick features in layers. Considering the CNN process, the earlier layers extract general features while the later layers extract detailed features. Therefore, the researchers feed 10 training images to Model (B), and the Grad-cam points out that the 18th layer extracts the nick features (refer to Figure 8). From this analysis, Model (B) properly extracts the nick features in the 18th layer.

6. Next Steps

This study focuses on developing a propeller nick detection model to improve maturity in the propeller nick detection inspection tasks. While the model is trained on images of closely resembling nicks using appropriate hyperparameters and its explainability is analyzed through Grad-cam, several aspects remain outside the scope of this research. Future research areas could focus on practical implementation, assessing how well this model can detect actual nicks. Additionally, studies could explore the benefits of ergonomic design of the model prototype, its usability and utility for inspectors, and its practical applicability. Further research could also assess measurable and tangible benefits, such as risk reduction.

7. Conclusion

Detecting nicks is a crucial inspection task because nicks initiate tip failure and do not show prefailure symptoms. To improve inspection outcomes, the researchers developed a propeller nick detection model using the YOLO algorithm. This model can serve as the inspectors’ third eye in the propeller nick inspection tasks. This model can detect nicks that inspectors may overlook during inspections. As a result, this study has the potential to enhance inspection accuracy and reduce general aviation accidents caused by propeller failure.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding

This research received no specific grant from any funding agency.

Open Research

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

References

1 Federal Aviation Administration, AC 20-37E-Aircraft Propeller Maintenance, 2005, Accessed 21 August 2023. Available at: https://www.faa.gov/regulations_policies/advisory_circulars/index.cfm/go/document.information/documentID/22576.
Google Scholar
2 Code of Federal Regulations, 14 CFR Part 65 Subpart D -- Mechanics, Accessed 21 August 2023b. Available at: https://www.ecfr.gov/current/title-14/part-65/subpart-D.
Google Scholar
3 Code of Federal Regulations, 14 CFR Part 43-- Maintenance, Preventive Maintenance, Rebuilding, and Alteration, Accessed 21 August 2023a. Available at: https://www.ecfr.gov/current/title-14/part-43.
Google Scholar
4 Brown S., Machine Learning, Explained, 2023, Accessed 21 August 2023. Available at: https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained.
Google Scholar
5 Hardesty L., Explained: Neural Networks, 2017, MIT News. Accessed 21 August 2023, https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414.
Google Scholar
6 Redmon J., Divvala S., Girshick R., and Farhadi A., You Only Look Once: Unified, Real-Time Object Detection, Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Presented at the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, IEEE, 779–788, https://doi.org/10.1109/CVPR.2016.91, 2-s2.0-84986308404.
10.1109/CVPR.2016.91
Google Scholar
7 Terven J., Córdova-Esparza D.-M., and Romero-González J.-A., A Comprehensive Review of Yolo Architectures in Computer Vision: From yolov1 to yolov8 and yolo-nas, Machine Learning and Knowledge Extraction. (2023) 5, no. 4, 1680–1716, https://doi.org/10.3390/make5040083.
10.3390/make5040083
Web of Science® Google Scholar
8 Jocher G., Chaurasia A., and Qiu J., YOLO by Ultralytics, 2023, Accessed 21 August 2023 https://github.com/ultralytics/ultralytics.
Google Scholar
9 Ju R.-Y. and Cai W., Fracture Detection in Pediatric Wrist Trauma X-ray Images Using YOLOv8 Algorithm, Scientific Reports. (2023) 13, no. 1, https://doi.org/10.1038/s41598-023-47460-7, 37973984.
10.1038/s41598-023-47460-7
PubMed Web of Science® Google Scholar
10 Phan Q. B. and Nguyen T. T., A Novel Approach for PV Cell Fault Detection Using YOLOv8 and Particle Swarm Optimization, Proceedings of the 2023 IEEE 66th International Midwest Symposium on Circuits and Systems (MWSCAS), 2023, IEEE, 634–638, https://doi.org/10.1109/MWSCAS57524.2023.10406139.
10.1109/MWSCAS57524.2023.10406139
Google Scholar
11 King R., Brief Summary of YOLOv8 Model Structure, 2023, Accessed 21 August 2023. https://github.com/ultralytics/ultralytics/issues/189.
Google Scholar
12 Wang C. Y., Mark Liao H. Y., Wu Y. H., Chen P. Y., Hsieh J. W., and Yeh I. H., CSPNet: A New Backbone That Can Enhance Learning Capability of CNN, Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Presented at the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020, IEEE, 1571–1580, https://doi.org/10.1109/CVPRW50498.2020.00203.
10.1109/CVPRW50498.2020.00203
Google Scholar
13 Lin T.-Y., Dollár P., Girshick R., He K., Hariharan B., and Belongie S., Feature Pyramid Networks for Object Detection, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, IEEE, 2117–2125.
10.1109/CVPR.2017.106
Google Scholar
14 Liu S., Qi L., Qin H., Shi J., and Jia J., Path Aggregation Network for Instance Segmentation, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, IEEE, 8759–8768.
10.1109/CVPR.2018.00913
Google Scholar
15 Aust J., Shankland S., Pons D., Mukundan R., and Mitrovic A., Automated Defect Detection and Decision-Support in Gas Turbine Blade Inspection, Aerospace. (2021) 8, no. 2, https://doi.org/10.3390/aerospace8020030.
10.3390/aerospace8020030
Web of Science® Google Scholar
16 Kim Y. H. and Lee J. R., Videoscope-Based Inspection of Turbofan Engine Blades Using Convolutional Neural Networks and Image Processing, Structural Health Monitoring. (2019) 18, no. 5–6, 2020–2039, https://doi.org/10.1177/1475921719830328, 2-s2.0-85062462056.
10.1177/1475921719830328
Web of Science® Google Scholar
17 Li X., Wang C., Ju H., and Li Z., Surface Defect Detection Model for Aero-Engine Components Based on Improved YOLOv5, Applied Sciences. (2022) 12, no. 14, https://doi.org/10.3390/app12147235.
10.1007/978-3-031-12762-5
Google Scholar
18 Mohtasham Khani M., Vahidnia S., Ghasemzadeh L., Ozturk Y. E., Yuvalaklioglu M., Akin S., and Ure N. K., Deep-Learning-Based Crack Detection With Applications for the Structural Health Monitoring of Gas Turbines, Structural Health Monitoring. (2020) 19, no. 5, 1440–1452, https://doi.org/10.1177/1475921719883202.
10.1177/1475921719883202
Google Scholar
19 Shang H., Sun C., Liu J., Chen X., and Yan R., Deep Learning-Based Borescope Image Processing for Aero-Engine Blade In-Situ Damage Detection, Aerospace Science and Technology. (2022) 123, 107473, https://doi.org/10.1016/j.ast.2022.107473.
10.1016/j.ast.2022.107473
Web of Science® Google Scholar
20 Shang H., Wu J., Sun C., Liu J., Chen X., and Yan R., Global Prior Transformer Network in Intelligent Borescope Inspection for Surface Damage Detection of Aeroengine Blade, IEEE Transactions on Industrial Informatics. (2023) 19, no. 8, 8865–8877, https://doi.org/10.1109/TII.2022.3222300.
10.1109/TII.2022.3222300
Web of Science® Google Scholar
21 Xiao R., Wan X., Li Y., Guan X., and Huang X., Semantic Damage Segmentation for Aircraft Engine Borescope Videos, Journal of Aerospace Information Systems. (2022) 19, no. 8, 565–573, https://doi.org/10.2514/1.I011051.
10.2514/1.I011051
Web of Science® Google Scholar
22 Ren I., An Ensemble Machine Vision System for Automated Detection of Surface Defects in Aircraft Propeller Blades, 2020, Georgia Institute of Technology.
Google Scholar
23 Ren I., Zahiri F., Sutton G., Kurfess T., and Saldana C., A Deep Ensemble Classifier for Surface Defect Detection in Aircraft Visual Inspection, Smart and Sustainable Manufacturing Systems. (2020) 4, no. 1, 81–94, https://doi.org/10.1520/SSMS20200031.
10.1520/SSMS20200031
Web of Science® Google Scholar
24 Li J., Su Z., Geng J., and Yin Y., Real-Time Detection of Steel Strip Surface Defects Based on Improved YOLO Detection Network, IFAC-PapersOnLine. (2018) 51, no. 21, 76–81, https://doi.org/10.1016/j.ifacol.2018.09.412, 2-s2.0-85054432094.
10.1016/j.ifacol.2018.09.412
Google Scholar
25 Deyin Z., Penghui W., Mingwei T., Conghan C., Li W., and Wenxuan H., Investigation of Aircraft Surface Defects Detection Based on YOLO Neural Network, Proceedings of the 2020 7th International Conference on Information Science and Control Engineering (ICISCE), 2020, IEEE, 781–785, https://doi.org/10.1109/ICISCE50968.2020.00165.
10.1109/ICISCE50968.2020.00165
Google Scholar
26 Zhang D., Zeng N., and Lin L., Detection of Blades Damages in Aero Engine, Proceedings of the 2020 Chinese Automation Congress (CAC), Presented at the 2020 Chinese Automation Congress (CAC), 2020, IEEE, 6129–6134, https://doi.org/10.1109/CAC51589.2020.9327202.
10.1109/CAC51589.2020.9327202
Google Scholar
27 Xiaoxun Z., Xinyu H., Xiaoxia G., Xing Y., Zixu X., Yu W., and Huaxin L., Research on Crack Detection Method of Wind Turbine Blade Based on a Deep Learning Method, Applied Energy. (2022) 328, 120241, https://doi.org/10.1016/j.apenergy.2022.120241.
10.1016/j.apenergy.2022.120241
Google Scholar
28 Zhang R. and Wen C., SOD-YOLO: A Small Target Defect Detection Algorithm for Wind Turbine Blades Based on Improved YOLOv5, Advanced Theory and Simulations. (2022) 5, no. 7, 2100631, https://doi.org/10.1002/adts.202100631.
10.1002/adts.202100631
Web of Science® Google Scholar
29 Caldwell F. W., Propellers for Aircraft Engines of High Power Output, Journal of the Aeronautical Sciences. (1937) 5, no. 2, 37–52, https://doi.org/10.2514/8.510.
10.2514/8.510
Google Scholar
30 Caldwell F. W., Aircraft-Propeller Development and Testing Summarized: Part 1, SAE Transactions. (1934) 29, 297–310.
Google Scholar
31 Caldwell F. W., Aircraft-Propeller Development and Testing Summarized: Part 2 (Concluded), SAE Transactions. (1934) 29, 349–358.
Google Scholar
32 Campbell G. and Lahey R., A Survey of Serious Aircraft Accidents Involving Fatigue Fracture, International Journal of Fatigue. (1984) 6, no. 1, 25–30, https://doi.org/10.1016/0142-1123(84)90005-7, 2-s2.0-0021192518.
10.1016/0142-1123(84)90005-7
Web of Science® Google Scholar
33 Kushan M. C., Diltemiz S. F., and Sackesen İ., Failure Analysis of an Aircraft Propeller, Engineering Failure Analysis. (2007) 14, no. 8, 1693–1700, https://doi.org/10.1016/j.engfailanal.2006.11.069, 2-s2.0-34250716028.
10.1016/j.engfailanal.2006.11.069
Web of Science® Google Scholar
34 Hobbs A., An Overview of Human Factors in Aviation Maintenance, 2008, Aviation Research and Analysis Report No. AR-2008-055, Australian Transport Safety Bureau.
Google Scholar
35 Keller J., Mendonca F. C., and Cutter J. E., Collegiate Aviation Pilots: Analyses of Fatigue Related Decision-Making Scenarios, International Journal of Aviation, Aeronautics, and Aerospace. (2019) 6, no. 4, https://doi.org/10.15394/ijaaa.2019.1360.
10.15394/ijaaa.2019.1360
Google Scholar
36 V7, The AI Data Engine for Computer Vision & Generative AI, Accessed 21 August 2023. https://www.v7labs.com/.
Google Scholar
37 Liashchynskyi P. and Liashchynskyi P., Grid Search, Random Search, Genetic Algorithm: A Big Comparison for NAS, 2019, https://arxiv.org/abs/1912.06059.
Google Scholar
38 Davis J. and Goadrich M., The Relationship Between Precision-Recall and ROC Curves, Proceedings of the 23rd International Conference on Machine Learning-ICML’06, Presented at the the 23rd International Conference, 2006, ACM Press, 233–240, https://doi.org/10.1145/1143844.1143874, 2-s2.0-34250727580.
10.1145/1143844.1143874
Google Scholar
39 Powers D. M. W., Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness and Correlation, 2020, https://arxiv.org/abs/2010.16061.
Google Scholar
40 Pedregosa F., Varoquaux G., Gramfort A., Michel V., Thirion B., Grisel O., Blondel M., Prettenhofer P., Weiss R., Dubourg V., and Vanderplas J., Scikit-learn: Machine Learning in Python, Journal of Machine Learning Research. (2011) 12, 2825–2830.
Web of Science® Google Scholar
41 Gildenblat J., Advanced AI Explainability for PyTorch, 2023, Accessed 20 September 2023. Available at: https://github.com/jacobgil/pytorch-grad-cam.
Google Scholar
42 Rigved and Zelenchuk A., Eigen CAM for YOLO V8 Interpretability, Accessed 20 September 2023. https://github.com/rigvedrs/YOLO-V8-CAM.
Google Scholar
43 Egi Y., Hajyzadeh M., and Eyceyurt E., Drone-Computer Communication Based Tomato Generative Organ Counting Model Using YOLO V5 and Deep-Sort, Agriculture. (2022) 12, no. 9, https://doi.org/10.3390/agriculture12091290.
10.3390/agriculture12091290
Google Scholar
44 Kartashov O. O., Chernov A. V., Alexandrov A. A., Polyanichenko D. S., Ierusalimov V. S., Petrov S. A., and Butakova M. A., Machine Learning and 3D Reconstruction of Materials Surface for Nondestructive Inspection, Sensors. (2022) 22, no. 16, https://doi.org/10.3390/s22166201, 36015958.
10.3390/s22166201
PubMed Web of Science® Google Scholar
45 Li J., Cong Y., Zhou L., Tian Z., and Qiu J., Super-Resolution-Based Part Collaboration Network for Vehicle Re-Identification, World Wide Web. (2023) 26, no. 2, 519–538, https://doi.org/10.1007/s11280-022-01060-z.
10.1007/s11280-022-01060-z
Web of Science® Google Scholar

All articles

Detectable Inspection: Propeller Nick Detection Model Development Using YOLOv8

Abstract

1. Introduction