SWaveNet: elastic-wave sequence-to-image learning for buried polyethylene pipeline localisation

Accurate localisation of buried polyethylene (PE) pipelines remains challenging because surface-acquired elastic-wave signals are weak, noisy, and sensitive to soil conditions. This study proposes SWaveNet, a Subsurface Waveform-to-Image Network, for elastic-wave sequence-to-image learning for buried PE pipeline localisation. The proposed framework directly maps multi-channel time-domain elastic-wave signals to two-dimensional pipeline localisation maps, reducing the dependence on explicit wave-velocity estimation and handcrafted inversion procedures. To enhance field signal quality, a hybrid denoising method combining singular spectrum analysis and weighted anisotropic total variation, termed SSA-WATV, is introduced to suppress noise while preserving reflection-related waveform features. A hybrid dataset based on numerical simulations and field measurements is constructed for model development and validation. Experimental results demonstrate that SWaveNet achieves high localisation accuracy and robustness under noisy conditions. Compared with MLP, CNN, and traditional time-domain superposition method, SWaveNet exhibits lower localisation errors, stronger noise robustness, and reduced directional bias. These results indicate that elastic-wave sequence-to-image learning provides an effective data-driven solution for automated localisation of buried PE pipelines and highlight its potential for intelligent non-destructive evaluation of underground non-metallic infrastructure.