State-of-the-art Machine Learning Advances in Reliability-based Design, Integrity Assessment, Inspection and Maintenance of Pipelines: A Systematic Review

Machine learning (ML) has increasingly been adopted in pipeline engineering, extending traditional physics-based, statistical, and empirical diagnostics toward data-driven frameworks. This review systematically examines ML applications in reliability-based design, structural integrity evaluation, condition monitoring, and inspection planning of pipelines. The literature shows a shift from conventional case-specific supervised learning toward transferable, hybrid, metaheuristic, and physics-informed techniques. These frameworks decompose signals, quantify uncertainty, use graph-based knowledge representation, and enhance generalizability through physics-informed learning, ranging from theory-guided features and architectures to limited forms of constraint enforcement. However, progress is limited by small, low-quality laboratory datasets, little focus on fatigue crack growth and multi-defect interactions, and heavy reliance on opaque surrogate models. A comparative analysis reveals persistent bottlenecks in generalization, interpretability, and the coupling of material, environmental, and operational variables. Three main research frontiers emerge: (1) large-scale, multi-source benchmark datasets, (2) physics-informed and interpretable learning frameworks that bridge domain mechanics and ML algorithms, and (3) standardized evaluation protocols and field-level validation schemes. The review concludes with a research roadmap to accelerate the move from algorithm-centric experimentation to tractable, trustworthy ML frameworks for pipeline integrity management. Overall, this study provides guidance for improving pipeline integrity to support safe and sustainable energy transport through resilient infrastructures.