| dc.description.abstract | The research in this thesis centers on the image data interpretation capabilities of autonomous underwater systems in the offshore oil and gas industry responsible for visual inspection and monitoring of underwater pipelines and detection of hazards on pipeline surfaces. The main contribution of research is a framework that provides the solutions to overcome the vulnerabilities of artificial intelligence methods during the underwater pipeline inspection by autonomous underwater systems through applied safety engineering.
Increased autonomy in autonomous underwater systems require a greater reliance on artificial intelligence technology for executing pipeline inspection tasks. However, the artificial intelligence for pipeline inspection is limited by several data interpretation challenges. The shortcomings of artificial intelligence for autonomous systems during offshore pipeline hazard inspection can result in catastrophic environmental damage and substantial financial losses for the oil and gas industry. Imbalanced and underrepresented data can cause the artificial intelligence methods, such as machine learning, anomaly detection, and computer vision, to form biases in favor of more represented data with atendency to reproduce biases learned from data. Underrepresented data can be disregarded as noise during anomaly detection due to their inclination toward efficiency and sacrificing anomalies as tolerable collateral damage. Current methods focus primarily on data content with no regard for the context behind data, yielding conclusions primarily based on correlation and not causation, further causing the occurrence of false alarms during anomaly detection that are a significant drawback during real-time operations. Furthermore, the acute lack of annotated training image data of offshore pipelines and lack of hazard evidence in the data results in the reliance on inexplainable, unsupervised methods. Therefore, one of the main contributions of this research is using the methods for risk and hazard analysis to semi-supervise anomaly detection methods and generate synthetic images of pipeline hazards for extrapolating and annotating the training data. Risk analysis aids anomaly detection in identify the types of anomalies that are recognized as risks, by analyzing low-probability, high-consequence event detection. Furthermore, due to the acute disorganization of categorization and definition of anomalies in current research, this research proposes a redefined anomaly categorization for autonomous underwater systems operations based on hazard behavior and traditional anomaly classification. Finally, this research examines the complex and connected properties of offshore pipeline inspections and offers future directions in rethinking the artificial intelligence methods for pipeline inspection with autonomous underwater systems. The general theme of this thesis lies in risk-informed approaches to address the fundamental challenge of finding early true anomalies and avoiding false alarms when no label exists to inform us of the anomaly or its properties by giving context to anomaly detection methods to comprehend data points not by their labels but by how they relate to one another. | en_US |
