A physics-constrained deep learning framework for feature recognition in long-distance natural gas pipelines

Abstract

The safety assurance of long-distance natural gas pipelines depends on accurate perception of their service condition, and this capability, in turn, relies on effective identification of features such as bending deformation, dents, and weld anomalies. However, when bending strain signals are used for feature recognition, existing methods still suffer from insufficient discrimination of composite pipeline features and limited classification accuracy. To address these issues, this paper proposes a physics-constrained deep learning framework for long-distance natural gas pipeline features. The framework introduces the Symmetric Temporal Convolutional Network and the Conformer-V module to extract high-dimensional spatiotemporal information from pipeline features, enhance the perception of bidirectional spatial context, and capture local geometric symmetry patterns. Taking the Euler–Bernoulli beam theory as an engineering prior approximation, the physics-constrained modulation module learns information related to local bending response through a curvature regression head and a physical consistency loss, and further maps the predicted curvature into gating weights to modulate the backbone features, thereby achieving collaborative feature recognition through data-driven learning and physical constraints. Validation, ablation, and comparative experiments conducted on a 200 km inertial measurement unit field measurement dataset from a long-distance natural gas pipeline show that the proposed method achieves a classification accuracy of 98.94%, demonstrating strong discriminative performance for composite pipeline features.