Axial-stress classification of steel pipes from magnetic remanence via topology-aware multi-probe graph learning: controlled full-scale validation

Axial stress from geohazards threatens pipeline integrity, yet in-line quantification remains challenging. Remanence-based magneto-mechanical sensing is promising, but many approaches treat multi-probe readings independently and underuse array structure. We propose a topology-aware graph learning framework for seven-level axial-stress classification from multi-channel magnetic signals, with an empirically constrained auxiliary loss (ECAL) that encourages magneto-mechanical consistency. The circumferential probe layout is encoded as graphs and processed by a multi-graph attention network with efficient channel attention (MGAT-ECA) to capture probe-to-probe coupling and cross-channel interactions. We evaluate the approach via controlled full-scale axial tensile tests under high-remanence (HR) and natural initial magnetisation, using group-aware cross-validation and a held-out test set. Under HR, the framework achieves 99.1% ± 0.7% accuracy across 0–120 MPa with 23.1 ± 2.4 ms inference per sample (93.1% ± 1.1% under the natural state). The model trained on a Q235 Φ508 mm pipe transfers to an independent Q235 Φ1016 mm pipe without retraining, achieving 84.1% accuracy. Ablations support the contributions of the remanent bias, differential multi-channel inputs and ECAL, while attention weights emphasise the axial channel and localised circumferential coupling. Finally, we outline compatibility with magnetic flux leakage workflows, leveraging upstream remanence for downstream stress screening.