Session: 08-05-01 AI‑Enabled Marine Sensing, Monitoring & Digital Twins

Submission Number: 182210

Deep Learning Framework for Pipeline Defect Characterization Using Distributed Strain Data 

Pipeline structural health monitoring requires not only the detection of damage but also reliable estimation of defect size and location. Distributed fiber optic sensing provides dense strain measurements along pipelines; however, converting these measurements into quantitative defect information remains challenging. This paper presents a digital twin based approach for characterizing wall-thinning defects using strain data along a helical sensing path in simulation.

Finite element model of a pressurized 8-inch steel pipe was developed, and strain data was extracted along a virtual helical fiber path with 300 sampling points. Internal wall-thinning defects with different lengths and depths were simulated to form a seed dataset. To reduce the computational cost, a physics-consistent data augmentation method was adopted. The augmentation strategy was validated against an independent set of finite element simulations and showed good agreement.

Convolutional neural network was trained for inverse defect characterization and compared with a gradient-boosting regression model. Validation using independent finite element data shows that the convolutional model provides more accurate and stable predictions of defect depth and length, as well as improved robustness in defect localization. The baseline model performs well only under augmented in-distribution conditions and shows reduced accuracy when evaluated on pure finite element data. The results indicate that spatial feature learning plays a key role in the inversion of helical strain signatures and supports the use of physics-informed digital twins for pipeline structural health monitoring.

Keywords:Structural Health Monitoring (SHM); Distributed Fiber Optic Sensing (DFOS); Digital Twin; Physics-Informed Deep Learning; Finite Element Analysis (FEA); Pipeline Integrity; Convolutional Neural Networks (CNN); Defect Characterization; XGBoost

Presenting Author: Xuejun Wang A*STAR Institute of High Performance Computing (A*STAR IHPC)

Presenting Author Biography: Dr. Xuejun Wang is a Principal Scientist at A*STAR’s Institute of High Performance Computing (IHPC). Holding a Ph.D. in Structural & Mechanics from NTU (2009). His research focuses on computational mechanics, structural health monitoring, and digital-twin-enabled predictive maintenance of critical infrastructure. He has led and co-investigated multiple R&D projects on pipeline integrity, coastal protection, integrating digital modeling with distributed sensing for real-time condition assessment.

Authors:

Xuejun Wang A*STAR Institute of High Performance Computing (A*STAR IHPC)
Zhuangjian Liu A*STAR Institute of High Performance Computing (A*STAR IHPC)