Abstract The integrity management of subsea pipelines requires manual, time-consuming, and subjective review of vast quantities of underwater imagery, creating significant operational bottlenecks and costs for asset operators. This report presents the successful development and validation of a Human-in-the-Loop (HITL) AI system designed to augment expert inspectors by automating the initial screening of this data. The core objective was to create a robust deep learning model to accurately triage images, deprioritising those depicting healthy assets and flagging potential anomalies for focused human review. The methodology centered on a transfer learning approach, fine- tuning a MobileNetV2 architecture on an expert-annotated dataset of 2,166 real-world inspection images. A key technical challenge was a severe class imbalance, with anomalies requiring review constituting only about 8% of the data. This was systematically addressed through class weighting, a two-phase training strategy, and the implementation of an EarlyStopping callback to prevent model overfitting and ensure robust generalisation. The final, optimised AI model demonstrated a significant success, proving its value as a powerful tool for inspectors by delivering substantial efficiency gains. The system's primary impact is its ability to drastically reduce the manual review burden. For high- consequence assets, a "Maximum Safety" mode successfully identified 100% of all true anomalies in the test set, guaranteeing no potential threat is missed. Even in this most cautious setting (Maximum Safety mode), the system still delivered a significant 30% reduction in manual labour, proving that safety and efficiency gains are not mutually exclusive. In a balanced configuration, the model reduces the required human workload by a 75%, while still correctly identifying over two-thirds (68%) of all true anomalies. Critically, the system's sensitivity can be precisely tuned to meet specific operational needs and risk levels. This proven ability to configure the AI's behaviour is a core success of the project. The successful outcome of this proof-of-concept establishes a clear, data-driven pathway to reducing inspection cycle times from weeks to hours, improving the consistency of assessments, reduce the risk of sustained attention fatigue leading to critical anomalies being missed, and enabling a proactive approach to pipeline integrity management, all while preserving essential human judgment for all final decisions.
Rahnama et al. (Mon,) studied this question.