Physics-guided graph Transformer for composite pipeline damage detection

Xuan Li; Yue Hu; Pugen Zhang; Jinsha Xu; Yi Chen; Haibin Miao; Xinguang Ren; Fucai Li

doi:10.37965/jdmd.2026.1675

Authors

Xuan Li School of Mechanical and Power Engineering, East China University of Science and Technology, Shanghai, China https://orcid.org/0009-0005-5533-5670
Yue Hu School of Mechanical and Power Engineering, East China University of Science and Technology, Shanghai, China https://orcid.org/0000-0002-8090-9112
Pugen Zhang Shanghai Institute of Special Equipment Inspection and Technical Research Co., Ltd, Shanghai, China https://orcid.org/0000-0002-6811-0340
Jinsha Xu Shanghai Institute of Special Equipment Inspection and Technical Research Co., Ltd, Shanghai, China https://orcid.org/0009-0003-4857-7853
Yi Chen Shanghai Institute of Special Equipment Inspection and Technical Research Co., Ltd, Shanghai, China
Haibin Miao PETROCHINA DAGANG PETROCHEMICAL COMPANY, Tianjin, China
Xinguang Ren PETROCHINA DAGANG PETROCHEMICAL COMPANY
Fucai Li State Key Laboratory of Mechanical System and Vibration, Shanghai Jiao Tong University, Shanghai, China

DOI:

https://doi.org/10.37965/jdmd.2026.1675

Abstract

Composite pipelines are widely used in oil and gas transportation owing to their high specific strength and corrosion resistance. However, guided-wave-based damage localization remains challenging due to multimodal wave propagation, noise interference, complex propagation-path relationships, and limited labeled damage locations. To address these challenges, this study proposes a physics-guided graph Transformer (PGGT) framework for composite pipeline damage localization.

In the proposed method, guided-wave signals are first preprocessed, and actuator–receiver propagation paths are represented as graph nodes. A physics-guided radius graph is constructed by incorporating the spatial configuration of piezoelectric transducers and guided-wave propagation characteristics, enabling multi-channel guided-wave signals to be mapped into a physically meaningful non-Euclidean graph structure. A convolutional module is employed to extract local temporal features from guided-wave responses, a graph attention module is utilized to model topology relationships among signal-path nodes, and a Transformer module is introduced to capture long-range global dependencies among propagation paths.

Comparative experiments are conducted under four training–testing split strategies with unseen damage locations. The proposed PGGT is compared with VGG16, ResNet18, ChebyNet, GCN, GAT, Transformer, and PGGDL in terms of mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean square error (RMSE). The results demonstrate that PGGT achieves superior localization accuracy and generalization capability under limited labeled data. In the most challenging case, PGGT achieves a MAPE of 0.035 (that is, 3.5%), a MAE of 5.38 mm, and an RMSE of 6.23 mm. The proposed method provides an effective physics-informed learning framework for guided-wave-based damage localization of composite pipelines.

Conflict of Interest Statement

The authors declare no conflicts of interest.