Physics informed multiscale convolutional attention network for intelligent recognition of oil and gas pipeline leakage in ground penetrating radar data

Oil and gas pipeline leakage poses substantial risks to safety and the environment. Ground-penetrating radar (GPR) offers unique advantages for early leakage detection, but current intelligent recognition methods face four bottlenecks: loss of reflection polarity, missing electromagnetic constraints, poor cross-view synergy across A/B/C tri-view data, and simulation-to-real distribution shift under data scarcity. This paper proposes PolMS-GPRNet, a physics-guided multi-scale convolutional attention network. The polarity-sensitive multi-scale attention module (PS-MSCA) explicitly preserves reflection sign information through parallel positive and negative polarity streams. The Maxwell-equation-constrained physics-guided attention module (MEC-Attn) embeds the two-dimensional Finite-Difference Time-Domain (FDTD) discretized electromagnetic wave propagation residual into the attention loss as a physics-informed regularization term. The tri-view cross-view synergistic fusion module (TS-CVF) achieves progressive alignment at physical quantity, geometry, and topology levels via a hierarchical cross-view Transformer. Experiments on a dual dataset comprising 12,000 gprMax forward-modeled B-scans (with 4,800 companion 3D C-scan volumes) and 2,400 sandbox field-test samples demonstrate that the proposed method achieves a macro-averaged F1 (mF1) of 0.899 ± 0.006 across the five foreground categories (six classes including background), surpassing the strongest baseline DA-YOLO by 11.1 percentage points, with mAP@[0.5:0.95] reaching 0.612 for a 6.4-percentage-point improvement over the same baseline. F1 scores for oil seepage and gas escape improve by 15.3 and 15.6 percentage points, respectively. The direct simulation-to-real transfer gap compresses from a baseline range of 0.218 to 0.314 down to 0.108, a reduction exceeding 50%. Single-sample inference latency of 76 ms meets the requirements of typical GPR field-scanning speeds under a batched inference pipeline. This study provides an end-to-end, physics-guided, cross-dimensionally collaborative, prototype-level solution with engineering deployment potential for intelligent leakage recognition in oil and gas pipelines.