Artificial Intelligence (AI) Powered Detection and Identification of Pressure Vessel External Damage – YOLO-based Application

Artificial Intelligence (AI) Powered Detection and Identification of Pressure Vessel External Damage – YOLO-based Application | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Artificial Intelligence (AI) Powered Detection and Identification of Pressure Vessel External Damage – YOLO-based Application E. Seddik, M. Hewieg, R. Afify This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4573646/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Regular visual inspections of pressure vessels by qualified inspectors are vital in the oil and gas industry to ensure vessel integrity, prevent catastrophic failures, and avoid the consequences of misdiagnosis. In the present paper, a deep learning (DL) model is proposed to perform visual inspection of pressure vessels using the You Only Look Once (YOLO) v8 model. Initially, binary classification is developed to diagnose whether the pressure vessel's exterior shell is in excellent condition or includes damage using a training dataset of 5000 real shell surface on-site images from the Abu Madi gas field of the PETROBEL Company, Egypt. An additional model was trained on the same dataset of binary model for multi-class identification of external damages (i.e. corrosion, painting damage, mechanical damage, or brittle fracture). Three models were compared; namely YOLO v5, YOLO v8, and YOLO v9. The highest performance model was YOLO v8. Its detection accuracy reached 93% and 91.5% for binary classification and multi-class classification models, respectively. Implementing this solution in the inspection process will result in a cost reduction since it decreases the need for scaffolding and trained inspectors. This study provides a valuable roadmap for future research on image processing-based pressure vessel damage detection. Physical sciences/Engineering Physical sciences/Engineering/Mechanical engineering Pressure vessel External visual inspection Damage mechanisms Steel defect detection YOLO v8 algorithm Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Pressure vessels are essential assets in the operations of oil and gas companies, fulfilling various roles such as storage, separation, and filtration. The implications of their failure go beyond environmental and safety issues, highlighting the importance of regular visual inspections to maintain operational reliability [ 1 ]. For instance, the explosion incident at Minaa El Ahmadi in Kuwait in 2000 serves as a notable example of the severe consequences resulting from the absence of regular inspection procedures in oil and gas facilities. The initial investigation highlighted a lack of inspection and maintenance of a condensate line. The total property damage was estimated at $ 412 million. It resulted in three crude units being damaged, two reformers being destroyed, and a fire that was extinguished about nine hours after the initial explosion. Tragically, five people lost their lives, and 50 others were injured as a result of the incident. Data from the period spanning 2018 to 2022 indicates a total of 14 fatalities from falls from height, 9 fatalities from falling objects, and 15 fatalities from high-pressure system-related incidents [ 2 ]. The inspection process plays a critical role in evaluating the condition of a vessel by identifying and monitoring factors that may lead to damage. The data obtained from these inspections facilitates proactive risk management by guiding decisions related to future repairs, replacements, and inspection strategies. Consistent inspections are essential for the early detection and mitigation of potential hazards to prevent significant process safety incidents, such as fires, exposure to hazardous materials, and environmental harm [ 3 ]. Moreover, visual inspection is considered a fundamental and dependable method of inspection [ 4 ]. In the conventional approach, experienced human engineers typically conduct visual inspections of these assets that are located in challenging or congested environments. As a result, there is a growing need to relieve human engineers from hazardous, falling objects, and high-pressure system explosions tasks even without proper safety [ 5 ]. Pressure vessel is often situated at a high altitude, which poses time-consuming, labor-intensive and a challenge for inspection workers to access it. For instance, scaffolding a structural leg can take over 336 hours to build up, and the scaffolding alone can cost roughly $ 130,000 USD [ 6 ]. Consequently, there is an urgent need to develop a detection method based on computer vision algorithms in the operation and maintenance of oil and gas pressure vessels. YOLO series utilizes a direct network approach to provide the location and category information of the target. This characteristic has earned recognition for its faster detection speed, making it an end-to-end target detection algorithm. Unlike Faster R-CNN [ 7 ], the YOLO series transforms the object identification problem into a regression problem and classify the categories based on object identification network. Within the domain of corrosion research, there has been a surge in the development of various algorithms focused on the application of machine learning techniques. These algorithms can be divided into two primary groups based on their reliance on visual corrosion inspection methods. The first group consists of corrosion detection methods that utilize low-level features, the process of corrosion detection heavily depends on feature extraction. Color is recognized as a fundamental and frequently employed component in computer vision tasks. Pascual and Ortiz [ 8 ] illustrated a code-word dictionary built from stacked histograms from the red, green, and blue color channels. They trained a classifier to identify corrosion cases. Pascual and Ortiz [ 9 ] used color information for corrosion detection. Pereira et al. [ 10 ] situated Hue-Saturation-Value (HSV) values corresponding to corrosion zones within the Hue-Saturation (HS) plane. They employed a classifier to detect corrosion instances across the HSV color space and utilized the characterization of corrosion shapes and sizes for the identification of pitting corrosion. Hoang and Tran [ 11 ] introduced a texture analysis approach for corrosion detection. Their methodology involved extracting 78 features from the corroded area utilizing image color, gray-level run lengths (GLRL), and gray-level co-occurrence matrix (GLCM). Subsequently, they employed a support vector machine (SVM) to establish a decision boundary for classifying corrosion images. Hoang [ 12 ] utilized texture analysis for the identification of pitting corrosion. He computed statistical assessments of color channels, GLCM, and local binary patterns to characterize the attributes of the metal surface, resulting in 93 texture features. He detected pitting corrosion through the application of SVM. Khaire and Dhanalakshmi [ 13 ] mentioned that traditional methodologies require a foundational comprehension of corrosion and its optimal attributes. They identified the optimal attributes of corrosion remains a challenging task. While the second group includes corrosion detection systems, it employs deep learning techniques. CNN-based methodologies facilitate autonomous feature identification by eliminating the prerequisite for prior knowledge. A plethora of recent studies have explored the efficacy of CNNs in achieving high-precision corrosion detection. Atha and Jahanshahi [ 14 ] enhanced the classification and localization capabilities of a CNN network by implementing the sliding window technique. Du et al. [ 15 ], proposed dual-parallel CNN architecture for the classification of corrosion levels. Apart from the aforementioned approaches, several studies have employed CNN-based object detection techniques to identify instances of corrosion. Cha et al. [ 16 ] trained a Faster R-CNN model with a dataset of 1737 images to detect instances of bolt and steel corrosion. Atha and Jahanshahi [ 17 ] evaluated the effectiveness of convolutional neural networks in corrosion detection. Those networks could learn essential categorization features that were traditionally crafted manually in earlier methodologies. Convolutional neural networks offer several advantages, including the elimination of the need for human intervention and prior knowledge in feature design. Li et al. [ 18 ] detected steel corrosion using modified YOLO architecture that integrated all convolutional layers. Their studies yielded satisfactory outcomes in context of corrosion categorization or detection. Bastian et al. [ 19 ] aimed at developing a technique for visually inspecting pipeline flanges. Through the utilization of an improved YOLO v3 network, they notably enhanced the accuracy of identifying flange images, while advancing the convergence speed during network training, as indicated by experimental findings. Their method enabled comprehensive analysis of flange images by leveraging intelligent inspection robots and surveillance cameras within the vicinity. It facilitated intelligent operation and maintenance practices. Zhao et al. [ 20 ] developed a model for detecting imperfections on steel surfaces. They introduced a novel detector named RDD-YOLO, built upon YOLO v5 framework. RDD-YOLO demonstrated superior performance in terms of both speed and accuracy in detecting defects on steel surfaces. Zhang et al. [ 21 ] presented a model for detecting surface defects in wind turbines. Their improved model demonstrated the capability to effectively and accurately perform real-time target identification tasks on embedded devices with minimal power consumption. Jia et al. [ 22 ] established a model framework for detecting metal surface corrosion. Their study resulted in the development of the corrosion-YOLO v5s model; a framework for identifying metal surface corrosion based on YOLO v5s model. The objective of this research is to introduce a methodology for the real-time detection of structural defects in external pressure vessels within oil and gas plants, focusing on various visible damage. The initial step involves distinguishing between undamaged and damaged external shell surfaces. Subsequently, the study places particular emphasis on identifying corrosion, mechanical damage, painting defects, brittle fractures, cracks, and other external damage mechanisms affecting these vessels. since the application of artificial intelligence approaches to detect external damage of pressure vessels has not been directly covered by the body of research stated in the literature. 2. Methodology In the field of computer vision, object detection assumes a vital role by enabling the identification and accurate localization of objects within image or video data. This process entails two essential stages according to Zhao et al. [ 23 ]; localization (present one or more objects within the captured data) and classification (assign each object to a specific class label). The application of object detection principles can be leveraged for the detection of external damage mechanisms in pressure vessels. This approach requires a three-steps process; namely dataset, train, and deploy. 2.1. Dataset Preparation and Labeling One of the major obstacles encountered when employing object detection for the identification of external damage in oil and gas pressure vessels is the scarcity of publicly available datasets. To overcome this limitation, the current study undertook the compilation of a novel dataset that is specifically tailored for detecting various external damage mechanisms in oil and gas pressure vessels. The dataset consists of 2,000 real on-site images captured from a shell surface of pressure vessels at the PETROBEL-Abu Madi Field, Egypt. These images were manually collected and annotated with bounding boxes and corresponding class labels for various types of external damage. 2.2. Model Training YOLO has emerged as a prominent object detection technique within the research community, leveraging convolutional neural networks for efficient object localization and classification [ 24 ]. The current paper proposes a model architecture that utilizes YOLO deep learning methodology for the detection of external damage on pressure vessels. To elevate the damage detection process and avoid relying solely on a single model and its results, this study has utilized three distinct models; YOLO v5, YOLO v8, and YOLO v9, in the training process for damage detection mechanisms. These versions, provided by Ultralytics, are the most recent and user-friendly models available. They are simplifying the training process and delivering quick and accurate results. The selection of these models is based on their suitability for real-time object detection applications. They give their ability to process and analyze data in a timely and precise manner. 2.3. Model Deployment Trained model is deployed for real-world application. This deployment allows for the automatic detection of external damage mechanisms in pressure vessel inspections, potentially improving safety and efficiency. In the next section, two models are discussed; Damage Detection Model (DDM) and Damage Identification Model (DIM). 3. Models 3.1 Damage Detection Model (DDM) This model is an AI-driven model that oversees determining whether there is damage along the pressure vessel's external shell as a binary classification (Good or Damage). It offers a valuable first-stage approach for pressure vessel integrity evaluation where: To enhance the model's ability to identify defects, it is crucial that it can recognize any deviation from the norm as a defect, regardless of the specific type of defect. By reducing the number of classes that the model is required to detect, the model's accuracy can be significantly improved. Smaller set of classes leads to more accurate detection and classification. 3.1.1. Dataset Collection and Annotation The present training dataset for the detection model consisted of 2000 real on-site (PETROBEL- Abu Madi Field) images of the shell surface. They are categorized using two distinct states, each case containing 1,000 images; i.e. good or damaged. The images were scaled to 640 pixels and saved in the jpg format. The dataset employed for this study was annotated by a certified Non-Destructive Testing (NDT) inspector with six years of specialized experience in oil and gas plant inspections by using Roboflow, an online free annotation platform, to manually label each image within the dataset. A range of image augmentation techniques were implemented to enhance the adaptability of the model for recognizing pressure vessel external damages and expanding the dataset. These techniques include vertical flipping, Rotation +/-90, static cropping, and saturation modification, as shown in Fig. 1 . In order to bolster the model's capacity for generalization in real-world contexts, a range of data augmentation strategies were implemented. These augmentation methodologies not only enriched the dataset, increasing the image count from 2000 to 5000, as illustrated in Table 1 , but also successfully alleviated the potential for overfitting and consequently improving the model's capacity for generalization. Following the augmentation process, the dataset was split into two categories: 85% was used for training and 15% was used for validation. In order to test the model's performance in actual testing, 200 images that were unique from the original dataset (the model had never been seen before) were used to test the model. The paid version of Roboflow can be used. However, the size of the current dataset will be considered sufficient for this study if the dataset has proven effective in achieving satisfactory results and accuracy rates in damage detection and identification tasks. The present study focuses on demonstrating the feasibility and effectiveness of the proposed methodology. The obtained results can serve as a basis for future scientific and industrial research, which may aim at expanding the data set and improving the accuracy and reliability of the proposed methodology. The present study trained three different YOLO versions models for object detection using pre-processed images with a size of 640x640 pixels. The training regimen employed a batch size of 16, spanned 100 epochs, and utilized an IoU threshold of 0.5. Google Colab provided the computational resources for faster training. This configuration facilitates the training and evaluation of the model, ensuring its performance and accuracy across various tasks. Table 1 Comparison between the dataset before and after augmentation. Dataset Original dataset After data augmentation Training dataset (85%) Validation dataset (15%) Percentage Testing dataset (unseen) Number of Images 2000 5000 4250 750 100% 200 Good 1000 2500 2125 375 50% 100 Damage 1000 2500 2125 375 50% 100 3.1.2. Model Evaluation Indicators To comprehensively evaluate the trained YOLO models' object detection capabilities, this study adopts three well-established metrics: precision, recall, and mean Average Precision mAP. This complete assessment technique allows for a more accurate knowledge of the model's detection performance, making it easier to select the best model for the task. Precision, Eq. (1), is a metric that determines the proportion of correctly identified positive instances (true positives) out of the total instances predicted as positive (true positives and false positives). It measures how accurate the positive predictions are: Precision= \(\frac{Tp}{Tp+Fp}\) (1) Where \(Tp\) : Number of true positive samples. \(Fp\) : Number of false positive samples. Figure 2 presents the precision curve for three models; namely YOLO v5, YOLO v8, and YOLO v9, which illustrates the relationship between the precision and the number of epochs. This curve illustrates how the precision of the models increases as the number of epochs increases after that it becomes almost constant. The precision curve provides a visual representation of the model's learning process. The curve shows that the precision for all classes (Damage and Good) is high with the highest value of 90% for YOLO v8 model. Recall, Eq. (2), is a metric that determines the proportion of correctly identified positive instances (true positives) out of the total instances (true positives and false negative). Recall = \(\frac{Tp}{Tp+Fn}\) (2) Where \(Tp\) : Number of true positive samples. \(Fn\) : Number of false negative samples. Recall Confidence Curve, Fig. 3 , shows the relationship between the recall and epochs using three YOLO models; YOLO v5, YOLO v8, and YOLO v9. In this curve, the x-axis represents the number of epochs, the y-axis represents the recall. The curve shows that YOLO v8 model has the highest recall of 83%. Mean Average Precision, Eq. ( 3 ), is a metric that measures how well the model performs in object detection tasks. $$mAP=\frac{\sum _{K=1}^{N}PR}{N}$$ 3 Where N: Number of detected sample categories. P: Precision R: Recall Figure 4 , illustrates the relationship between mAP and the number of epochs to compare the overall detection performance of YOLO v5, YOLO v8, and YOLO v9 models. YOLO v8 model achieves the highest mAP of 86.9%. Consequently, it is a superior model that consistently makes accurate detections. 3.1.3. Model training evaluation Table 2 summarize the results’ values of the three previous indicators for the three used models; YOLO v5, YOLO v8, and YOLO v9. YOLO v8 model demonstrated remarkable mAP performance within the same framework, conversely, YOLO v5 and YOLO v9 exhibited lower values. Despite YOLO v9 being the most recent model released in February 2024, YOLO v8 model not only fulfills real-time detection requirements but also enhances detection accuracy and offers increased adaptability and practical applicability. Table 2 . Comparative analysis of different models: performance results Model Precision (%) Recall (%) mAP (%) YOLO v5 87.99 82.40 84.94 YOLO v8 90.09 82.67 86.92 YOLO v9 88.07 80.06 84.29 Confusion matrix is used to visualize the model performance. It provides insights into how many predictions are correct (True Positives (TP) and True Negatives (TN)) and how many are incorrect (false positives and false negatives). This helped identify the strengths and weaknesses of the model and improve its performance. The onfusion matrix shows that the model is performing well in the damage and good classes, with TP = 0.88 and 0.82, respectively. 3.2. Damage Identification Model (DIM) This model is an AI-driven model that recognizes the type of damage on the pressure vessel's external shell, including corrosion, brittle fracture, mechanical damage, and painting damage, as well as the good condition as multi-class classification. 3.2.1. Dataset Collection and Annotation Leveraging a six-year-experienced NDT inspector for manual labeling and image augmentation via Roboflow enhanced the dataset's diversity and real-world applicability for pressure vessel damage detection. These techniques encompassed vertical flipping, rotations of +/- 90 degrees, static cropping of image regions, and saturation, as shown in Fig. 5 . The same used dataset of the external shell of the pressure vessel were utilized, with 50% of the images depicting the pressure vessel in good condition, 30% depicting evidence of corrosion, 17% depicting brittle fracture, 2% mechanical damage showing, and 1% showing signs of Painting Damage, as shown in Table 3 . As mentioned before, image augmentation not only expanded the dataset from 2,000 to 5,000 samples but also incorporated real-world variations, mitigating overfitting and enhancing generalizability. Table 3 Comparison of the dataset before and after augmentation. Dataset Original dataset After data augmentation Training dataset (85%) Validation dataset (15%) Percentage Testing dataset (unseen) Number of Images 2000 5000 4250 750 100% 200 Good 1000 2500 2125 375 50% 40 Corrosion 600 1500 1274 226 30% 40 Brittle Fracture 330 850 723 127 17% 40 Mechanical Damage 45 100 85 15 2% 40 Painting Damage 25 50 43 7 1% 40 All images within the dataset were pre-split with 85% (4250 images) and 15% (750 images) training-validation splits and 100 epochs of training on Google Colab to optimize its learning model. 3.2.2. Model Evaluation Indicators To comprehensively evaluate the YOLO models’ object detection, this study employs Precision, recall, and mAP, and enables a robust assessment for optimal model selection. Figure 6 , depicts the Precision Confidence of three models; YOLO v5, YOLO v8, and YOLO v9. It shows that YOLO v8 model's Precision (correctly identified positive cases) reaching a peak of 90.3%, representing its effectiveness in accurate object detection through various damage classes (Good, Corrosion, Brittle Fracture, Mechanical Damage, Painting Damage). Figure 7 depicts the Recall Confidence Curve. It shows the relationship between recall and epochs for three models; YOLO v5, YOLO v8, and YOLO v9. It demonstrates that the model has a high recall value of 78% during training with YOLO v8 model. Figure 8 concisely compares mAP metrics of three YOLO object detection models; YOLO v5, YOLO v8, and YOLO v9. YOLO v8 achieved the highest mAP value of 79%, indicating its superior ability to detect objects accurately and consistently. 3.2.3. Model training evaluation This model employed three different models; YOLO v5, YOLO v8, and YOLO v9, on an identical dataset and training / testing procedures. The results, shown in Table 4 , show that YOLO v8 overrides both YOLO v5 and YOLO v9 and demonstrate its advantage in real-time object detection accuracy and practical applicability. Table 4 Comparative analysis of different models: performance results. Model Precision (%) Recall (%) mAP (%) YOLO v5 78.2 78.2 77.2 YOLO v8 90.3 78.7 79.1 YOLO v9 85.4 73.1 77.5 Confusion matrix is a valuable analytical tool that enables the estimation of key performance metrics. It provides a detailed breakdown of YOLO v8 model's strengths and weaknesses. It shows a strong performance in the Corrosion (0.88 true positive rate), Good (0.84), Mechanical Damage (0.82), and Brittle Fracture (0.80) classes, and a low performance in the Painting Damage (0. 41) class. 4. Test 4.1. Damage Detection Model (DDM) Test dataset for the Damage Detection Model (DDM) contained 200 unseen images with size 640x640 pixels and JPG format. DDM was equally tested in two classes with 100 images each. Images were labeled as “Damage” if the pressure vessel external shell was affected by any kind of damage, while those in good condition were labeled as “Good”. This study assessed the applicability of YOLO v5, YOLO v8, and YOLO v9 for real-time binary classification of external damage states in oil and gas pressure vessels. The performance of pre-trained weights associated with each model was evaluated to identify the most accurate and efficient solution for this specific task. An assessment was conducted to compare the performance of YOLO v5, YOLO v8, and YOLO v9 in detecting external damage mechanisms on oil and gas pressure vessels. YOLO v8 model achieved a superior detection rate of (93%) as it correctly classified 186 images of the test set and (7%) remained undetected of 14 images, as shown in Table 5 . However, no conflict occurred where there are no misclassified images. Table 5 Performance evaluation of object detection models. YOLO v5 178 22 89% YOLO v8 186 14 93% YOLO v9 181 19 90.5% 4.2. Damage Identification Model (DIM) The Damage Identification Model (DIM) test dataset was never seen before by the DIM. The dataset consisted of 640x640 pixel JPG images. Five classes represented the output: Good, Corrosion, Brittle Fracture, Mechanical Damage, or Painting Damage. Each of the 5 classes was represented by 40 images. Pressure vessels with no flaws at all were labeled as “Good”, while those with damage were labeled according to the corresponding classes. This study investigated the efficacy of YOLO v5, YOLO v8, and YOLO v9 for real-time multiclass classification of external damage mechanisms in oil and gas pressure vessels. Pre-trained weights for each model were evaluated to identify the most accurate and computationally efficient solution for this application. Among the evaluated models, YOLO v8 achieved the highest detection rate of (91.5%); 183 images with all identified images classified correctly and 17 images of the test set images remained undetected. However, no conflict happens where there are no misclassified images, as shown in Table 6 . Table 6 Performance evaluation of object detection models. Model No. of correctly detected No. of not detected Accuracy YOLO v5 175 25 87.5% YOLO v8 183 17 91.5% YOLO v9 179 21 89.5 5. Conclusion Artificial Intelligence (AI) was used in this research to detect and identify external damages on pressure vessel exterior shells. This solution was developed for the oil and gas industry to avoid possible financial impacts of pressure vessel failures. This study proposed three based deep learning models; namely YOLO v5, YOLO v8, and YOLO v9, to determine whether the pressure vessel external shell is in good condition or suffers from any form of damage. Real pressure vessel images were collected on-site by the second author at Abu Madi gas field of PETROBEL Company, Egypt. The collected data was structured into an AI-friendly dataset which was then labeled manually to perform supervised learning. The dataset was split into training, validation, and testing samples. Two distinct computer vision models were developed within this framework: Damage Detection Model ( DDM ): This binary classifier differentiated between pressure vessels in "Good" condition and those exhibiting any form of external damage. DDM achieved an impressive average detection rate of 93% by the highest performance model YOLO v8. Damage Identification Model (DIM ): This model focuses on identifying the specific type of damage existed on a pressure vessel. DIM distinguished between "Good" condition and four damage categories; namely Corrosion, Brittle Fracture, Mechanical Damage, and Painting Damage. DIM demonstrated a noteworthy average identification accuracy of 91.5% by YOLO v8. Limitations and Future Work This research demonstrates the significant potential of deep learning for automating pressure vessel inspections, ultimately contributing to more efficient and reliable maintenance practices within the oil and gas industry. Future advancements in this field could encompass: Dataset Expansion : Enriching the training data with diverse pressure vessel types and damage scenarios. Real-Time Integration : Integrating the model with real-time inspection robots for on-site defect detection. Transfer Learning : Adapting the model for defect detection in other industrial applications beyond oil and gas. This research covers the way for a robust AI-based pressure vessel inspection system, potentially changing safety and cost-efficiency in oil and gas and offering broader industrial applications. The next research milestone aims to collect a similar dataset of pressure vessels to evaluate the performance of the model. A remotely operated vehicle (ROV) is currently in process to perform photography tasks outside and inside of pressure vessels to serve the data collection process. Declarations Author Contribution Conceptualization, E.S., M.H. and R.A.; methodology, E.S. and M.H.; software, E.S. and M.H.; validation, E.S. and M.H.; formal analysis, E.S. and M.H.; investigation, E.S., M.H. and R.A.; re-sources, M.H.; data curation, M.H.; writing—original draft preparation, M.H.; writing—review and editing, E.S., M.H. and R.A.; supervision, E.S. and R.A. Data Availability The datasets generated and/or analysed during the current study are not publicly available due [The data set contains images representing the intellectual property of the inspection sector, Petrobel company, Egypt. Public release is likely to compromise the confidentiality of inspections or inspected equipment.] but are available from the corresponding author on reasonable request. References Leijian, Y., Yang, E., Ren, P., Luo, C., Dobie, G., Gu, D., and Yan, X., “Inspection Robots in Oil and Gas Industry: A Review of Current Solutions and Future Trends”, International Conference on Automation and Computing (ICAC), Lancaster University, Lancaster UK, (2019). International Association of Oil & Gas Producers, “IOGP Annual Report”, (2023). American Petroleum Institute, “Pressure Vessel Inspection Code: In-Service Inspection, Rating, Repair, and Alteration Downstream Segment API 510” (9th ed.), Washington, DC: American Petroleum Institute, (2006). Ivan, K. Klarák, J. Bulej,V. Sága, M. Kandera, M. Hajdučík, A. and Tucki, K., “Approach to Automated Visual Inspection of Objects Based on Artificial Intelligence”, Applied Sciences, vol. 12, pp. 864, (2022). Orozhiyathumana, A. Lopes, S. Chalil, S. Kapil, M. Piratla, K. and Gramopadhye, A., “A survey of automation-enabled human-in-the-loop systems for infrastructure visual inspection”, Automation in Construction, vol. 97, pp. 52–76, (2019). V. S. Kumar, “Innovative use of drones improves safety on asset integrity,” Abu Dhabi International Petroleum Exhibition & Conference, (2017). Shaoqing, R. He, K. Girshick, R. and Sun, J., “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, pp. 1137–1149, (2017). Pascual, F.B. and Ortiz, A., “Detection of Cracks and Corrosion for Automated Vessels Visual Inspection”, Development Frontiers in Artificial Intelligence and Applications, vol. 220, pp. 111–120, (2010). Pascual, F.B., and Ortiz, A., “Corrosion Detection for Automated Visual Inspection”, IntechOpen, Developments in corrosion protection, pp. 619–632, (2014). Pereira, M., Silva, J., Acciari, H., Codaro, E., and Hein, R., “Morphology Characterization and Kinetics Evaluation of Pitting Corrosion of Commercially Pure Aluminium by Digital Image Analysis”, Materials Sciences and Applications, vol.03, pp. 287–293, (2012). Hoang, N., and Tran, V., “Image Processing-Based Detection of Pipe Corrosion Using Texture Analysis and Metaheuristic-Optimized Machine Learning Approach”, Computational Intelligence and Neuroscience, vol. 2019, pp. 1–13, (2019). Hoang, N., “Image Processing-Based Pitting Corrosion Detection Using Metaheuristic Optimized Multilevel Image Thresholding and Machine-Learning Approaches”, Mathematical Problems in Engineering, vol.2020, pp.1–19, (2020). Khaire, U., and Dhanalakshmi, R., “Stability of Feature Selection Algorithm: A Review”, Journal of King Saud University - Computer and Information Sciences, vol.34, pp. 1060–1073, (2019). Atha, D., and Jahanshahi, M., “Evaluation of Deep Learning Approaches Based on Convolutional Neural Networks for Corrosion Detection” Structural Health Monitoring, vol.17, pp.1110–1128, (2017). Du, J., Yan, L., Wang, H. and Huang, Q., “Research on Grounding Grid Corrosion Classification Method Based on Convolutional Neural Network” MATEC Web Conferences, College of Electrical and Control Engineering, Xi’an University of Science and Technology, Xi’an, China, (2018). Cha, Y., Choi, W., Suh, G., Mahmoudkhani, S., Büyüköztürk, O., “Autonomous Structural Visual Inspection Using Region-Based Deep Learning for Detecting Multiple Damage Types.”, Computer-Aided Civil and Infrastructure Engineering, vol. 33, pp. 731–747, (2017). Atha, D., Jahanshahi, M., “Evaluation of deep learning approaches based on convolutional neural networks for corrosion detection”, Structural Health Monitoring: An International Journal, vol.17, pp. 1110–1128, (2017). Li, J., Su, Z., Geng,J., and Yin Y., “Real-time Detection of Steel Strip Surface Defects Based on Improved YOLO Detection Network” IFAC-Papers OnLine, vol.51, pp. 76–81, (2018). Bastian B., Jaspreeth, N., Ranjith S., and Jiji, C., “Visual inspection and characterization of external corrosion in pipelines using deep neural network”, NDT & E International, vol.107, pp. 102–134, (2019). Zhao, C., Shu, X., Yan, X., Zuo, X., and Zhu, F., “RDD-YOLO: A modified YOLO for detection of steel surface defects. Measurement”, vol.214, pp.112–776, (2023). Zhang, Y., Yang, Y., Sun, J., Ji,R., Zhang, P., and Shan, H., "Surface defect detection of wind turbine based on lightweight YOLOv5s model", Measurement, vol.220, pp. 113–222, (2023). Jia, Z., Fu, M., Zhao, X., and Cui, Z., “Intelligent identification of metal corrosion based on Corrosion-YOLOv5s”, Displays, vol.76, pp. 102–367, (2023). Zhao, Z., Zheng, P. Xu, S., and Wu X. “Object Detection With Deep Learning: A Review”, IEEE Transactions on Neural Networks and Learning Systems, vol.30, pp.3212–3232, (2019). Redmon, J., Santosh, D., Ross, G. and Ali, F., “You Only Look Once: Unified, Real-Time Object Detection” IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, (2016). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4573646","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":320903050,"identity":"555efbe5-130f-46bc-b234-4918c2f950eb","order_by":0,"name":"E. Seddik","email":"","orcid":"","institution":"Arab Academy for Science, Technology and Maritime Transport","correspondingAuthor":false,"prefix":"","firstName":"E.","middleName":"","lastName":"Seddik","suffix":""},{"id":320903052,"identity":"d4d4b50f-5ddf-45b8-9e83-dd68517155fe","order_by":1,"name":"M. Hewieg","email":"data:image/png;base64,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","orcid":"","institution":"Arab Academy for Science, Technology and Maritime Transport","correspondingAuthor":true,"prefix":"","firstName":"M.","middleName":"","lastName":"Hewieg","suffix":""},{"id":320903054,"identity":"37d18baf-75f7-47ce-8b29-1ed05dcc4aae","order_by":2,"name":"R. Afify","email":"","orcid":"","institution":"Arab Academy for Science, Technology and Maritime Transport","correspondingAuthor":false,"prefix":"","firstName":"R.","middleName":"","lastName":"Afify","suffix":""}],"badges":[],"createdAt":"2024-06-13 05:25:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4573646/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4573646/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":60339143,"identity":"60db4379-a28c-4e01-9e1d-4d6f6257e59d","added_by":"auto","created_at":"2024-07-15 17:52:47","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":324253,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/aa04fd50c236b240539313f9.jpg"},{"id":60339144,"identity":"3d0dea2b-7bb8-4369-bce7-dddc9a826874","added_by":"auto","created_at":"2024-07-15 17:52:47","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":96457,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/9979498e5c2753910e6a5ff8.jpg"},{"id":60338101,"identity":"bb95672f-79c0-4599-b8b4-42594d25b8af","added_by":"auto","created_at":"2024-07-15 17:44:47","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":93688,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/8d4522bd5d6461ed54530223.jpg"},{"id":60338103,"identity":"d53cc51b-79d5-432d-a377-21701aa5dca5","added_by":"auto","created_at":"2024-07-15 17:44:47","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":89257,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/8ebdf934fe5d0fc847809adc.jpg"},{"id":60338107,"identity":"ad48f584-a9f1-4f6a-bea4-aadcd21f6fbc","added_by":"auto","created_at":"2024-07-15 17:44:47","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":380242,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/eac47d1e533ef48422b1c07f.jpg"},{"id":60338108,"identity":"3dbc0bba-75b2-40bb-b3a8-02c8a09e3a95","added_by":"auto","created_at":"2024-07-15 17:44:47","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":123066,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/177d59148c294996094b565e.jpg"},{"id":60338105,"identity":"93da8c7a-4451-4015-a3b9-8854f64bd064","added_by":"auto","created_at":"2024-07-15 17:44:47","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":116833,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/fe22b185c5b8a334f2a23228.jpg"},{"id":60339145,"identity":"a9800e06-b540-4882-9da4-a3c2efe2d95d","added_by":"auto","created_at":"2024-07-15 17:52:47","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":104303,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"Picture8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/8497abb93a3e715aa9d0340a.jpg"},{"id":65366083,"identity":"62eba7f4-883d-4ffb-b7e9-bd0d9d1569bb","added_by":"auto","created_at":"2024-09-26 14:09:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2036069,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4573646/v1/47aa5076-d403-4519-9a58-40b2dd48a499.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial Intelligence (AI) Powered Detection and Identification of Pressure Vessel External Damage – YOLO-based Application","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003ePressure vessels are essential assets in the operations of oil and gas companies, fulfilling various roles such as storage, separation, and filtration. The implications of their failure go beyond environmental and safety issues, highlighting the importance of regular visual inspections to maintain operational reliability [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. For instance, the explosion incident at Minaa El Ahmadi in Kuwait in 2000 serves as a notable example of the severe consequences resulting from the absence of regular inspection procedures in oil and gas facilities. The initial investigation highlighted a lack of inspection and maintenance of a condensate line. The total property damage was estimated at \u003cspan\u003e$\u003c/span\u003e412\u0026nbsp;million. It resulted in three crude units being damaged, two reformers being destroyed, and a fire that was extinguished about nine hours after the initial explosion. Tragically, five people lost their lives, and 50 others were injured as a result of the incident. Data from the period spanning 2018 to 2022 indicates a total of 14 fatalities from falls from height, 9 fatalities from falling objects, and 15 fatalities from high-pressure system-related incidents [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe inspection process plays a critical role in evaluating the condition of a vessel by identifying and monitoring factors that may lead to damage. The data obtained from these inspections facilitates proactive risk management by guiding decisions related to future repairs, replacements, and inspection strategies. Consistent inspections are essential for the early detection and mitigation of potential hazards to prevent significant process safety incidents, such as fires, exposure to hazardous materials, and environmental harm [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Moreover, visual inspection is considered a fundamental and dependable method of inspection [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. In the conventional approach, experienced human engineers typically conduct visual inspections of these assets that are located in challenging or congested environments. As a result, there is a growing need to relieve human engineers from hazardous, falling objects, and high-pressure system explosions tasks even without proper safety [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003ePressure vessel is often situated at a high altitude, which poses time-consuming, labor-intensive and a challenge for inspection workers to access it. For instance, scaffolding a structural leg can take over 336 hours to build up, and the scaffolding alone can cost roughly \u003cspan\u003e$\u003c/span\u003e130,000 USD [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Consequently, there is an urgent need to develop a detection method based on computer vision algorithms in the operation and maintenance of oil and gas pressure vessels. YOLO series utilizes a direct network approach to provide the location and category information of the target. This characteristic has earned recognition for its faster detection speed, making it an end-to-end target detection algorithm. Unlike Faster R-CNN [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], the YOLO series transforms the object identification problem into a regression problem and classify the categories based on object identification network.\u003c/p\u003e \u003cp\u003eWithin the domain of corrosion research, there has been a surge in the development of various algorithms focused on the application of machine learning techniques. These algorithms can be divided into two primary groups based on their reliance on visual corrosion inspection methods. The first group consists of corrosion detection methods that utilize low-level features, the process of corrosion detection heavily depends on feature extraction. Color is recognized as a fundamental and frequently employed component in computer vision tasks. \u003cb\u003ePascual and Ortiz\u003c/b\u003e [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] illustrated a code-word dictionary built from stacked histograms from the red, green, and blue color channels. They trained a classifier to identify corrosion cases. \u003cb\u003ePascual and Ortiz\u003c/b\u003e [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] used color information for corrosion detection. \u003cb\u003ePereira et al.\u003c/b\u003e [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] situated Hue-Saturation-Value (HSV) values corresponding to corrosion zones within the Hue-Saturation (HS) plane. They employed a classifier to detect corrosion instances across the HSV color space and utilized the characterization of corrosion shapes and sizes for the identification of pitting corrosion. \u003cb\u003eHoang and Tran\u003c/b\u003e [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] introduced a texture analysis approach for corrosion detection. Their methodology involved extracting 78 features from the corroded area utilizing image color, gray-level run lengths (GLRL), and gray-level co-occurrence matrix (GLCM). Subsequently, they employed a support vector machine (SVM) to establish a decision boundary for classifying corrosion images. \u003cb\u003eHoang\u003c/b\u003e [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] utilized texture analysis for the identification of pitting corrosion. He computed statistical assessments of color channels, GLCM, and local binary patterns to characterize the attributes of the metal surface, resulting in 93 texture features. He detected pitting corrosion through the application of SVM. \u003cb\u003eKhaire and Dhanalakshmi\u003c/b\u003e [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] mentioned that traditional methodologies require a foundational comprehension of corrosion and its optimal attributes. They identified the optimal attributes of corrosion remains a challenging task.\u003c/p\u003e \u003cp\u003eWhile the second group includes corrosion detection systems, it employs deep learning techniques. CNN-based methodologies facilitate autonomous feature identification by eliminating the prerequisite for prior knowledge. A plethora of recent studies have explored the efficacy of CNNs in achieving high-precision corrosion detection. \u003cb\u003eAtha and Jahanshahi\u003c/b\u003e [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] enhanced the classification and localization capabilities of a CNN network by implementing the sliding window technique. \u003cb\u003eDu et al.\u003c/b\u003e [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], proposed dual-parallel CNN architecture for the classification of corrosion levels. Apart from the aforementioned approaches, several studies have employed CNN-based object detection techniques to identify instances of corrosion. \u003cb\u003eCha et al.\u003c/b\u003e [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] trained a Faster R-CNN model with a dataset of 1737 images to detect instances of bolt and steel corrosion. \u003cb\u003eAtha and Jahanshahi\u003c/b\u003e [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] evaluated the effectiveness of convolutional neural networks in corrosion detection. Those networks could learn essential categorization features that were traditionally crafted manually in earlier methodologies. Convolutional neural networks offer several advantages, including the elimination of the need for human intervention and prior knowledge in feature design. \u003cb\u003eLi et al.\u003c/b\u003e [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] detected steel corrosion using modified YOLO architecture that integrated all convolutional layers. Their studies yielded satisfactory outcomes in context of corrosion categorization or detection. \u003cb\u003eBastian et al.\u003c/b\u003e [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] aimed at developing a technique for visually inspecting pipeline flanges. Through the utilization of an improved YOLO v3 network, they notably enhanced the accuracy of identifying flange images, while advancing the convergence speed during network training, as indicated by experimental findings. Their method enabled comprehensive analysis of flange images by leveraging intelligent inspection robots and surveillance cameras within the vicinity. It facilitated intelligent operation and maintenance practices. \u003cb\u003eZhao et al.\u003c/b\u003e [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] developed a model for detecting imperfections on steel surfaces. They introduced a novel detector named RDD-YOLO, built upon YOLO v5 framework. RDD-YOLO demonstrated superior performance in terms of both speed and accuracy in detecting defects on steel surfaces. \u003cb\u003eZhang et al.\u003c/b\u003e [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] presented a model for detecting surface defects in wind turbines. Their improved model demonstrated the capability to effectively and accurately perform real-time target identification tasks on embedded devices with minimal power consumption. \u003cb\u003eJia et al.\u003c/b\u003e [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] established a model framework for detecting metal surface corrosion. Their study resulted in the development of the corrosion-YOLO v5s model; a framework for identifying metal surface corrosion based on YOLO v5s model.\u003c/p\u003e \u003cp\u003eThe objective of this research is to introduce a methodology for the real-time detection of structural defects in external pressure vessels within oil and gas plants, focusing on various visible damage. The initial step involves distinguishing between undamaged and damaged external shell surfaces. Subsequently, the study places particular emphasis on identifying corrosion, mechanical damage, painting defects, brittle fractures, cracks, and other external damage mechanisms affecting these vessels. since the application of artificial intelligence approaches to detect external damage of pressure vessels has not been directly covered by the body of research stated in the literature.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eIn the field of computer vision, object detection assumes a vital role by enabling the identification and accurate localization of objects within image or video data. This process entails two essential stages according to \u003cb\u003eZhao et al.\u003c/b\u003e [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]; localization (present one or more objects within the captured data) and classification (assign each object to a specific class label). The application of object detection principles can be leveraged for the detection of external damage mechanisms in pressure vessels. This approach requires a three-steps process; namely dataset, train, and deploy.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Dataset Preparation and Labeling\u003c/h2\u003e \u003cp\u003eOne of the major obstacles encountered when employing object detection for the identification of external damage in oil and gas pressure vessels is the scarcity of publicly available datasets. To overcome this limitation, the current study undertook the compilation of a novel dataset that is specifically tailored for detecting various external damage mechanisms in oil and gas pressure vessels. The dataset consists of 2,000 real on-site images captured from a shell surface of pressure vessels at the PETROBEL-Abu Madi Field, Egypt. These images were manually collected and annotated with bounding boxes and corresponding class labels for various types of external damage.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Model Training\u003c/h2\u003e \u003cp\u003eYOLO has emerged as a prominent object detection technique within the research community, leveraging convolutional neural networks for efficient object localization and classification [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The current paper proposes a model architecture that utilizes YOLO deep learning methodology for the detection of external damage on pressure vessels. To elevate the damage detection process and avoid relying solely on a single model and its results, this study has utilized three distinct models; YOLO v5, YOLO v8, and YOLO v9, in the training process for damage detection mechanisms. These versions, provided by Ultralytics, are the most recent and user-friendly models available. They are simplifying the training process and delivering quick and accurate results. The selection of these models is based on their suitability for real-time object detection applications. They give their ability to process and analyze data in a timely and precise manner.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Model Deployment\u003c/h2\u003e \u003cp\u003eTrained model is deployed for real-world application. This deployment allows for the automatic detection of external damage mechanisms in pressure vessel inspections, potentially improving safety and efficiency. In the next section, two models are discussed; Damage Detection Model (DDM) and Damage Identification Model (DIM).\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Models","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Damage Detection Model (DDM)\u003c/h2\u003e\n \u003cp\u003eThis model is an AI-driven model that oversees determining whether there is damage along the pressure vessel\u0026apos;s external shell as a binary classification (Good or Damage). It offers a valuable first-stage approach for pressure vessel integrity evaluation where:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eTo enhance the model\u0026apos;s ability to identify defects, it is crucial that it can recognize any deviation from the norm as a defect, regardless of the specific type of defect.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003eBy reducing the number of classes that the model is required to detect, the model\u0026apos;s accuracy can be significantly improved. Smaller set of classes leads to more accurate detection and classification.\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e\n \u003ch2\u003e3.1.1. Dataset Collection and Annotation\u003c/h2\u003e\n \u003cp\u003eThe present training dataset for the detection model consisted of 2000 real on-site (PETROBEL- Abu Madi Field) images of the shell surface. They are categorized using two distinct states, each case containing 1,000 images; i.e. good or damaged.\u003c/p\u003e\n \u003cp\u003eThe images were scaled to 640 pixels and saved in the jpg format. The dataset employed for this study was annotated by a certified Non-Destructive Testing (NDT) inspector with six years of specialized experience in oil and gas plant inspections by using Roboflow, an online free annotation platform, to manually label each image within the dataset. A range of image augmentation techniques were implemented to enhance the adaptability of the model for recognizing pressure vessel external damages and expanding the dataset. These techniques include vertical flipping, Rotation +/-90, static cropping, and saturation modification, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eIn order to bolster the model\u0026apos;s capacity for generalization in real-world contexts, a range of data augmentation strategies were implemented. These augmentation methodologies not only enriched the dataset, increasing the image count from 2000 to 5000, as illustrated in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, but also successfully alleviated the potential for overfitting and consequently improving the model\u0026apos;s capacity for generalization. Following the augmentation process, the dataset was split into two categories: 85% was used for training and 15% was used for validation. In order to test the model\u0026apos;s performance in actual testing, 200 images that were unique from the original dataset (the model had never been seen before) were used to test the model. The paid version of Roboflow can be used. However, the size of the current dataset will be considered sufficient for this study if the dataset has proven effective in achieving satisfactory results and accuracy rates in damage detection and identification tasks.\u003c/p\u003e\n \u003cp\u003eThe present study focuses on demonstrating the feasibility and effectiveness of the proposed methodology. The obtained results can serve as a basis for future scientific and industrial research, which may aim at expanding the data set and improving the accuracy and reliability of the proposed methodology.\u003c/p\u003e\n \u003cp\u003eThe present study trained three different YOLO versions models for object detection using pre-processed images with a size of 640x640 pixels. The training regimen employed a batch size of 16, spanned 100 epochs, and utilized an IoU threshold of 0.5. Google Colab provided the computational resources for faster training. This configuration facilitates the training and evaluation of the model, ensuring its performance and accuracy across various tasks.\u0026nbsp;\u003c/p\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison between the dataset before and after augmentation.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOriginal dataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAfter data augmentation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining dataset\u003c/p\u003e\n \u003cp\u003e(85%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eValidation dataset\u003c/p\u003e\n \u003cp\u003e(15%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePercentage\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTesting dataset\u003c/p\u003e\n \u003cp\u003e(unseen)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of Images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4250\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGood\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e375\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDamage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e375\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\n \u003ch2\u003e3.1.2. Model Evaluation Indicators\u003c/h2\u003e\n \u003cp\u003eTo comprehensively evaluate the trained YOLO models\u0026apos; object detection capabilities, this study adopts three well-established metrics: precision, recall, and mean Average Precision mAP. This complete assessment technique allows for a more accurate knowledge of the model\u0026apos;s detection performance, making it easier to select the best model for the task.\u003c/p\u003e\n \u003cp\u003ePrecision, Eq.\u0026nbsp;(1), is a metric that determines the proportion of correctly identified positive instances (true positives) out of the total instances predicted as positive (true positives and false positives). It measures how accurate the positive predictions are:\u003c/p\u003e\n \u003cp\u003ePrecision=\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\frac{Tp}{Tp+Fp}\\)\u003c/span\u003e\u003c/span\u003e (1)\u003c/p\u003e\n \u003cp\u003eWhere\u003c/p\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u0026nbsp;\u003cspan class=\"mathinline\"\u003e\\(Tp\\)\u003c/span\u003e\u0026nbsp;\u003c/span\u003e: Number of true positive samples.\u003c/p\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u0026nbsp;\u003cspan class=\"mathinline\"\u003e\\(Fp\\)\u003c/span\u003e\u0026nbsp;\u003c/span\u003e: Number of false positive samples.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e presents the precision curve for three models; namely YOLO v5, YOLO v8, and YOLO v9, which illustrates the relationship between the precision and the number of epochs. This curve illustrates how the precision of the models increases as the number of epochs increases after that it becomes almost constant. The precision curve provides a visual representation of the model\u0026apos;s learning process. The curve shows that the precision for all classes (Damage and Good) is high with the highest value of 90% for YOLO v8 model.\u003c/p\u003e\n \u003cp\u003eRecall, Eq.\u0026nbsp;(2), is a metric that determines the proportion of correctly identified positive instances (true positives) out of the total instances (true positives and false negative).\u003c/p\u003e\n \u003cp\u003eRecall =\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\frac{Tp}{Tp+Fn}\\)\u003c/span\u003e\u003c/span\u003e (2)\u003c/p\u003e\n \u003cp\u003eWhere\u003c/p\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u0026nbsp;\u003cspan class=\"mathinline\"\u003e\\(Tp\\)\u003c/span\u003e\u0026nbsp;\u003c/span\u003e: Number of true positive samples.\u003c/p\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u0026nbsp;\u003cspan class=\"mathinline\"\u003e\\(Fn\\)\u003c/span\u003e\u0026nbsp;\u003c/span\u003e: Number of false negative samples.\u003c/p\u003e\n \u003cp\u003eRecall Confidence Curve, Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, shows the relationship between the recall and epochs using three YOLO models; YOLO v5, YOLO v8, and YOLO v9. In this curve, the x-axis represents the number of epochs, the y-axis represents the recall. The curve shows that YOLO v8 model has the highest recall of 83%.\u003c/p\u003e\n \u003cp\u003eMean Average Precision, Eq.\u0026nbsp;(\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e), is a metric that measures how well the model performs in object detection tasks.\u003c/p\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$$mAP=\\frac{\\sum _{K=1}^{N}PR}{N}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eWhere\u003c/p\u003e\n \u003cp\u003eN: Number of detected sample categories.\u003c/p\u003e\n \u003cp\u003eP: Precision\u003c/p\u003e\n \u003cp\u003eR: Recall\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, illustrates the relationship between mAP and the number of epochs to compare the overall detection performance of YOLO v5, YOLO v8, and YOLO v9 models. YOLO v8 model achieves the highest mAP of 86.9%. Consequently, it is a superior model that consistently makes accurate detections.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\n \u003ch2\u003e3.1.3. Model training evaluation\u003c/h2\u003e\n \u003cp\u003eTable 2 summarize the results\u0026rsquo; values of the three previous indicators for the three used models; YOLO v5, YOLO v8, and YOLO v9. YOLO v8 model demonstrated remarkable mAP performance within the same framework, conversely, YOLO v5 and YOLO v9 exhibited lower values. Despite YOLO v9 being the most recent model released in February 2024, YOLO v8 model not only fulfills real-time detection requirements but also enhances detection accuracy and offers increased adaptability and practical applicability.\u003c/p\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. Comparative analysis of different models: performance results\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"612\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"24.50980392156863%\" valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"27.77777777777778%\" valign=\"top\"\u003e\n \u003cp\u003ePrecision (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.65359477124183%\" valign=\"top\"\u003e\n \u003cp\u003eRecall (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"22.058823529411764%\" valign=\"top\"\u003e\n \u003cp\u003emAP (%)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"24.50980392156863%\" valign=\"top\"\u003e\n \u003cp\u003eYOLO v5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"27.77777777777778%\" valign=\"top\"\u003e\n \u003cp\u003e87.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.65359477124183%\" valign=\"top\"\u003e\n \u003cp\u003e82.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"22.058823529411764%\" valign=\"top\"\u003e\n \u003cp\u003e84.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"24.50980392156863%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eYOLO v8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"27.77777777777778%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e90.09\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.65359477124183%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e82.67\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"22.058823529411764%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e86.92\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"24.50980392156863%\" valign=\"top\"\u003e\n \u003cp\u003eYOLO v9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"27.77777777777778%\" valign=\"top\"\u003e\n \u003cp\u003e88.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.65359477124183%\" valign=\"top\"\u003e\n \u003cp\u003e80.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"22.058823529411764%\" valign=\"top\"\u003e\n \u003cp\u003e84.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003eConfusion matrix is used to visualize the model performance. It provides insights into how many predictions are correct (True Positives (TP) and True Negatives (TN)) and how many are incorrect (false positives and false negatives). This helped identify the strengths and weaknesses of the model and improve its performance. The onfusion matrix shows that the model is performing well in the damage and good classes, with TP\u0026thinsp;=\u0026thinsp;0.88 and 0.82, respectively.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2. Damage Identification Model (DIM)\u003c/h2\u003e\n \u003cp\u003eThis model is an AI-driven model that recognizes the type of damage on the pressure vessel\u0026apos;s external shell, including corrosion, brittle fracture, mechanical damage, and painting damage, as well as the good condition as multi-class classification.\u003c/p\u003e\n \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\n \u003ch2\u003e3.2.1. Dataset Collection and Annotation\u003c/h2\u003e\n \u003cp\u003eLeveraging a six-year-experienced NDT inspector for manual labeling and image augmentation via Roboflow enhanced the dataset\u0026apos;s diversity and real-world applicability for pressure vessel damage detection. These techniques encompassed vertical flipping, rotations of +/- 90 degrees, static cropping of image regions, and saturation, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eThe same used dataset of the external shell of the pressure vessel were utilized, with 50% of the images depicting the pressure vessel in good condition, 30% depicting evidence of corrosion, 17% depicting brittle fracture, 2% mechanical damage showing, and 1% showing signs of Painting Damage, as shown in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. As mentioned before, image augmentation not only expanded the dataset from 2,000 to 5,000 samples but also incorporated real-world variations, mitigating overfitting and enhancing generalizability.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of the dataset before and after augmentation.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOriginal dataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAfter data augmentation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining dataset\u003c/p\u003e\n \u003cp\u003e(85%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eValidation dataset\u003c/p\u003e\n \u003cp\u003e(15%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePercentage\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTesting dataset\u003c/p\u003e\n \u003cp\u003e(unseen)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of Images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4250\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGood\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e375\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCorrosion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e600\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1274\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e226\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBrittle Fracture\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e330\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e850\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e723\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e127\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e17%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMechanical Damage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePainting Damage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eAll images within the dataset were pre-split with 85% (4250 images) and 15% (750 images) training-validation splits and 100 epochs of training on Google Colab to optimize its learning model.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\n \u003ch2\u003e3.2.2. Model Evaluation Indicators\u003c/h2\u003e\n \u003cp\u003eTo comprehensively evaluate the YOLO models\u0026rsquo; object detection, this study employs Precision, recall, and mAP, and enables a robust assessment for optimal model selection. Figure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, depicts the Precision Confidence of three models; YOLO v5, YOLO v8, and YOLO v9. It shows that YOLO v8 model\u0026apos;s Precision (correctly identified positive cases) reaching a peak of 90.3%, representing its effectiveness in accurate object detection through various damage classes (Good, Corrosion, Brittle Fracture, Mechanical Damage, Painting Damage).\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e depicts the Recall Confidence Curve. It shows the relationship between recall and epochs for three models; YOLO v5, YOLO v8, and YOLO v9. It demonstrates that the model has a high recall value of 78% during training with YOLO v8 model.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e concisely compares mAP metrics of three YOLO object detection models; YOLO v5, YOLO v8, and YOLO v9. YOLO v8 achieved the highest mAP value of 79%, indicating its superior ability to detect objects accurately and consistently.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\n \u003ch2\u003e3.2.3. Model training evaluation\u003c/h2\u003e\n \u003cp\u003eThis model employed three different models; YOLO v5, YOLO v8, and YOLO v9, on an identical dataset and training / testing procedures. The results, shown in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, show that YOLO v8 overrides both YOLO v5 and YOLO v9 and demonstrate its advantage in real-time object detection accuracy and practical applicability.\u003c/p\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparative analysis of different models: performance results.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003emAP (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYOLO v5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e78.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e78.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e77.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eYOLO v8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e90.3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e78.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e79.1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYOLO v9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e73.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e77.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003eConfusion matrix is a valuable analytical tool that enables the estimation of key performance metrics. It provides a detailed breakdown of YOLO v8 model\u0026apos;s strengths and weaknesses. It shows a strong performance in the Corrosion (0.88 true positive rate), Good (0.84), Mechanical Damage (0.82), and Brittle Fracture (0.80) classes, and a low performance in the Painting Damage (0. 41) class.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4. Test","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Damage Detection Model (DDM)\u003c/h2\u003e \u003cp\u003eTest dataset for the Damage Detection Model (DDM) contained 200 unseen images with size 640x640 pixels and JPG format. DDM was equally tested in two classes with 100 images each. Images were labeled as \u0026ldquo;Damage\u0026rdquo; if the pressure vessel external shell was affected by any kind of damage, while those in good condition were labeled as \u0026ldquo;Good\u0026rdquo;.\u003c/p\u003e \u003cp\u003eThis study assessed the applicability of YOLO v5, YOLO v8, and YOLO v9 for real-time binary classification of external damage states in oil and gas pressure vessels. The performance of pre-trained weights associated with each model was evaluated to identify the most accurate and efficient solution for this specific task. An assessment was conducted to compare the performance of YOLO v5, YOLO v8, and YOLO v9 in detecting external damage mechanisms on oil and gas pressure vessels. YOLO v8 model achieved a superior detection rate of (93%) as it correctly classified 186 images of the test set and (7%) remained undetected of 14 images, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. However, no conflict occurred where there are no misclassified images.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance evaluation of object detection models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO v5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e178\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYOLO v8\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e186\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e14\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e93%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO v9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e181\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e90.5%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Damage Identification Model (DIM)\u003c/h2\u003e \u003cp\u003eThe Damage Identification Model (DIM) test dataset was never seen before by the DIM. The dataset consisted of 640x640 pixel JPG images. Five classes represented the output: Good, Corrosion, Brittle Fracture, Mechanical Damage, or Painting Damage. Each of the 5 classes was represented by 40 images. Pressure vessels with no flaws at all were labeled as \u0026ldquo;Good\u0026rdquo;, while those with damage were labeled according to the corresponding classes.\u003c/p\u003e \u003cp\u003eThis study investigated the efficacy of YOLO v5, YOLO v8, and YOLO v9 for real-time multiclass classification of external damage mechanisms in oil and gas pressure vessels. Pre-trained weights for each model were evaluated to identify the most accurate and computationally efficient solution for this application.\u003c/p\u003e \u003cp\u003eAmong the evaluated models, YOLO v8 achieved the highest detection rate of (91.5%); 183 images with all identified images classified correctly and 17 images of the test set images remained undetected. However, no conflict happens where there are no misclassified images, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance evaluation of object detection models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo. of correctly detected\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo. of not detected\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO v5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e175\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e87.5%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYOLO v8\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e183\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e17\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e91.5%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO v9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e179\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eArtificial Intelligence (AI) was used in this research to detect and identify external damages on pressure vessel exterior shells. This solution was developed for the oil and gas industry to avoid possible financial impacts of pressure vessel failures. This study proposed three based deep learning models; namely YOLO v5, YOLO v8, and YOLO v9, to determine whether the pressure vessel external shell is in good condition or suffers from any form of damage. Real pressure vessel images were collected on-site by the second author at Abu Madi gas field of PETROBEL Company, Egypt. The collected data was structured into an AI-friendly dataset which was then labeled manually to perform supervised learning. The dataset was split into training, validation, and testing samples. Two distinct computer vision models were developed within this framework:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDamage Detection Model\u003c/b\u003e (\u003cb\u003eDDM\u003c/b\u003e): This binary classifier differentiated between pressure vessels in \"Good\" condition and those exhibiting any form of external damage. DDM achieved an impressive average detection rate of \u003cb\u003e93%\u003c/b\u003e by the highest performance model YOLO v8.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDamage Identification Model (DIM\u003c/b\u003e): This model focuses on identifying the specific type of damage existed on a pressure vessel. DIM distinguished between \"Good\" condition and four damage categories; namely Corrosion, Brittle Fracture, Mechanical Damage, and Painting Damage. DIM demonstrated a noteworthy average identification accuracy of \u003cb\u003e91.5%\u003c/b\u003e by YOLO v8.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eLimitations and Future Work\u003c/b\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis research demonstrates the significant potential of deep learning for automating pressure vessel inspections, ultimately contributing to more efficient and reliable maintenance practices within the oil and gas industry. Future advancements in this field could encompass:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDataset Expansion\u003c/b\u003e: Enriching the training data with diverse pressure vessel types and damage scenarios.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eReal-Time Integration\u003c/b\u003e: Integrating the model with real-time inspection robots for on-site defect detection.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eTransfer Learning\u003c/b\u003e: Adapting the model for defect detection in other industrial applications beyond oil and gas.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThis research covers the way for a robust AI-based pressure vessel inspection system, potentially changing safety and cost-efficiency in oil and gas and offering broader industrial applications. The next research milestone aims to collect a similar dataset of pressure vessels to evaluate the performance of the model.\u003c/p\u003e \u003cp\u003eA remotely operated vehicle (ROV) is currently in process to perform photography tasks outside and inside of pressure vessels to serve the data collection process.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization, E.S., M.H. and R.A.; methodology, E.S. and M.H.; software, E.S. and M.H.; validation, E.S. and M.H.; formal analysis, E.S. and M.H.; investigation, E.S., M.H. and R.A.; re-sources, M.H.; data curation, M.H.; writing\u0026mdash;original draft preparation, M.H.; writing\u0026mdash;review and editing, E.S., M.H. and R.A.; supervision,\u0026nbsp;E.S.\u0026nbsp;and\u0026nbsp;R.A.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and/or analysed during the current study are not publicly available due [The data set contains images representing the intellectual property of the inspection sector, Petrobel company, Egypt. Public release is likely to compromise the confidentiality of inspections or inspected equipment.] but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLeijian, Y., Yang, E., Ren, P., Luo, C., Dobie, G., Gu, D., and Yan, X., \u0026ldquo;Inspection Robots in Oil and Gas Industry: A Review of Current Solutions and Future Trends\u0026rdquo;, International Conference on Automation and Computing (ICAC), Lancaster University, Lancaster UK, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInternational Association of Oil \u0026amp; Gas Producers, \u0026ldquo;IOGP Annual Report\u0026rdquo;, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmerican Petroleum Institute, \u0026ldquo;Pressure Vessel Inspection Code: In-Service Inspection, Rating, Repair, and Alteration Downstream Segment API 510\u0026rdquo; (9th ed.), Washington, DC: American Petroleum Institute, (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIvan, K. Klar\u0026aacute;k, J. Bulej,V. S\u0026aacute;ga, M. Kandera, M. Hajduč\u0026iacute;k, A. and Tucki, K., \u0026ldquo;Approach to Automated Visual Inspection of Objects Based on Artificial Intelligence\u0026rdquo;, Applied Sciences, vol. 12, pp. 864, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrozhiyathumana, A. Lopes, S. Chalil, S. Kapil, M. Piratla, K. and Gramopadhye, A., \u0026ldquo;A survey of automation-enabled human-in-the-loop systems for infrastructure visual inspection\u0026rdquo;, Automation in Construction, vol. 97, pp. 52\u0026ndash;76, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eV. S. Kumar, \u0026ldquo;Innovative use of drones improves safety on asset integrity,\u0026rdquo; Abu Dhabi International Petroleum Exhibition \u0026amp; Conference, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaoqing, R. He, K. Girshick, R. and Sun, J., \u0026ldquo;Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks\u0026rdquo;, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, pp. 1137\u0026ndash;1149, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePascual, F.B. and Ortiz, A., \u0026ldquo;Detection of Cracks and Corrosion for Automated Vessels Visual Inspection\u0026rdquo;, Development Frontiers in Artificial Intelligence and Applications, vol. 220, pp. 111\u0026ndash;120, (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePascual, F.B., and Ortiz, A., \u0026ldquo;Corrosion Detection for Automated Visual Inspection\u0026rdquo;, IntechOpen, Developments in corrosion protection, pp. 619\u0026ndash;632, (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePereira, M., Silva, J., Acciari, H., Codaro, E., and Hein, R., \u0026ldquo;Morphology Characterization and Kinetics Evaluation of Pitting Corrosion of Commercially Pure Aluminium by Digital Image Analysis\u0026rdquo;, Materials Sciences and Applications, vol.03, pp. 287\u0026ndash;293, (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoang, N., and Tran, V., \u0026ldquo;Image Processing-Based Detection of Pipe Corrosion Using Texture Analysis and Metaheuristic-Optimized Machine Learning Approach\u0026rdquo;, Computational Intelligence and Neuroscience, vol. 2019, pp. 1\u0026ndash;13, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoang, N., \u0026ldquo;Image Processing-Based Pitting Corrosion Detection Using Metaheuristic Optimized Multilevel Image Thresholding and Machine-Learning Approaches\u0026rdquo;, Mathematical Problems in Engineering, vol.2020, pp.1\u0026ndash;19, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhaire, U., and Dhanalakshmi, R., \u0026ldquo;Stability of Feature Selection Algorithm: A Review\u0026rdquo;, Journal of King Saud University - Computer and Information Sciences, vol.34, pp. 1060\u0026ndash;1073, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAtha, D., and Jahanshahi, M., \u0026ldquo;Evaluation of Deep Learning Approaches Based on Convolutional Neural Networks for Corrosion Detection\u0026rdquo; Structural Health Monitoring, vol.17, pp.1110\u0026ndash;1128, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu, J., Yan, L., Wang, H. and Huang, Q., \u0026ldquo;Research on Grounding Grid Corrosion Classification Method Based on Convolutional Neural Network\u0026rdquo; MATEC Web Conferences, College of Electrical and Control Engineering, Xi\u0026rsquo;an University of Science and Technology, Xi\u0026rsquo;an, China, (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCha, Y., Choi, W., Suh, G., Mahmoudkhani, S., B\u0026uuml;y\u0026uuml;k\u0026ouml;zt\u0026uuml;rk, O., \u0026ldquo;Autonomous Structural Visual Inspection Using Region-Based Deep Learning for Detecting Multiple Damage Types.\u0026rdquo;, Computer-Aided Civil and Infrastructure Engineering, vol. 33, pp.\u0026lt;direction:rtl;vertical-align:super;\u0026gt; \u0026lt;/direction:rtl;vertical-align:super;\u0026gt;731\u0026ndash;747, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAtha, D., Jahanshahi, M., \u0026ldquo;Evaluation of deep learning approaches based on convolutional neural networks for corrosion detection\u0026rdquo;, Structural Health Monitoring: An International Journal, vol.17, pp. 1110\u0026ndash;1128, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, J., Su, Z., Geng,J., and Yin Y., \u0026ldquo;Real-time Detection of Steel Strip Surface Defects Based on Improved YOLO Detection Network\u0026rdquo; IFAC-Papers OnLine, vol.51, pp. 76\u0026ndash;81, (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBastian B., Jaspreeth, N., Ranjith S., and Jiji, C., \u0026ldquo;Visual inspection and characterization of external corrosion in pipelines using deep neural network\u0026rdquo;, NDT \u0026amp; E International, vol.107, pp. 102\u0026ndash;134, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, C., Shu, X., Yan, X., Zuo, X., and Zhu, F., \u0026ldquo;RDD-YOLO: A modified YOLO for detection of steel surface defects. Measurement\u0026rdquo;, vol.214, pp.112\u0026ndash;776, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Y., Yang, Y., Sun, J., Ji,R., Zhang, P., and Shan, H., \"Surface defect detection of wind turbine based on lightweight YOLOv5s model\", Measurement, vol.220, pp. 113\u0026ndash;222, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJia, Z., Fu, M., Zhao, X., and Cui, Z., \u0026ldquo;Intelligent identification of metal corrosion based on Corrosion-YOLOv5s\u0026rdquo;, Displays, vol.76, pp. 102\u0026ndash;367, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, Z., Zheng, P. Xu, S., and Wu X. \u0026ldquo;Object Detection With Deep Learning: A Review\u0026rdquo;, IEEE Transactions on Neural Networks and Learning Systems, vol.30, pp.3212\u0026ndash;3232, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRedmon, J., Santosh, D., Ross, G. and Ali, F., \u0026ldquo;You Only Look Once: Unified, Real-Time Object Detection\u0026rdquo; IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, (2016).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Pressure vessel, External visual inspection, Damage mechanisms, Steel defect detection, YOLO v8 algorithm","lastPublishedDoi":"10.21203/rs.3.rs-4573646/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4573646/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eRegular visual inspections of pressure vessels by qualified inspectors are vital in the oil and gas industry to ensure vessel integrity, prevent catastrophic failures, and avoid the consequences of misdiagnosis. In the present paper, a deep learning (DL) model is proposed to perform visual inspection of pressure vessels using the You Only Look Once (YOLO) v8 model. Initially, binary classification is developed to diagnose whether the pressure vessel's exterior shell is in excellent condition or includes damage using a training dataset of 5000 real shell surface on-site images from the Abu Madi gas field of the PETROBEL Company, Egypt. An additional model was trained on the same dataset of binary model for multi-class identification of external damages (i.e. corrosion, painting damage, mechanical damage, or brittle fracture). Three models were compared; namely YOLO v5, YOLO v8, and YOLO v9. The highest performance model was YOLO v8. Its detection accuracy reached 93% and 91.5% for binary classification and multi-class classification models, respectively. Implementing this solution in the inspection process will result in a cost reduction since it decreases the need for scaffolding and trained inspectors. This study provides a valuable roadmap for future research on image processing-based pressure vessel damage detection.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence (AI) Powered Detection and Identification of Pressure Vessel External Damage – YOLO-based Application","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-07-15 17:44:42","doi":"10.21203/rs.3.rs-4573646/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"22d3bd45-1847-40d0-a7e6-9ec227c68954","owner":[],"postedDate":"July 15th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":33924262,"name":"Physical sciences/Engineering"},{"id":33924263,"name":"Physical sciences/Engineering/Mechanical engineering"}],"tags":[],"updatedAt":"2024-09-26T14:08:57+00:00","versionOfRecord":[],"versionCreatedAt":"2024-07-15 17:44:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4573646","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4573646","identity":"rs-4573646","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}