AI-based Detection of Pressure Vessel Internal Damage Mechanisms

Preserving the human factor while mitigating risks associated with confined spaces, cutting down on the overall time of the inspection process, and optimizing regular equipment maintenance costs are the three most important factors for the success of the inspection process in oil and gas fields. Due to cost concerns and to ensure safety, the pressure vessel inspection process requires alternatives to traditional labor-intensive visual inspection techniques. This paper proposes a deep learning (DL) model as an intelligent and accurate tool for visual inspection of pressure vessels using a training dataset of 5000 real internal shell surface on-site images from the Abu Madi gas field of the PETROBEL Company, Egypt. A six-year-experienced Non-Destructive Testing (NDT) inspector was leveraged for manual labeling of the dataset. The present detection model utilizes You Only Look Once (YOLO) v8 model from v5, v8 and v10 models to diagnose whether the pressure vessel's inner surfaces of shell is in good condition or includes damage. The developed YOLO v8 multi-class identification model successfully detects the state of the pressure vessel's inner shell; good condition, or in the existence of damage, corrosion, pitting corrosion, mechanical damage, or brittle fracture. After training the model on the used on-site dataset images, the test process reveals a detection accuracy up to 93.3%. During the test process, if a single image contains good and damaged parts or two different types of damage, the model can differentiate between those cases. Applying this model in the inspection process will reduce costs by excluding the need for scaffolding and qualified inspectors and reducing the number of injuries and fatalities recorded due to confined spaces.