CSAInvNet: A Multinetwork Collaborative Deep Learning Framework for GPR-Based Joint 3-D Inversion Imaging of Coal Seam Anomalies

摘要

Three-dimensional electromagnetic inversion and representation of coal seam anomalies based on ground-penetrating radar (GPR) data is a highly promising approach for real-time coal seam modeling, enabling safe, precise, and efficient sensing in coal mining shearer robots. However, traditional methods face challenges in handling sparse B-scan measurements and noise-contaminated signals while maintaining computational efficiency. To overcome these challenges, we introduce CSAInvNet, a novel collaborative learning framework that integrates three innovative neural networks for robust 3-D permittivity reconstruction. The proposed method first employs a transformer-enhanced interpolation network that generates high-density B-scan volumes from sparse inputs by learning cross-channel correlations. Then, an improved mask-guided CycleGAN architecture is proposed for simultaneously performing signal denoising and realistic noise synthesis through adversarial training with physical constraints. Finally, a 3-D inverter is established for the inverse mapping from processed B-scan data volumes to anomaly distributions with geometric preservation. Extensive simulation and field experiments demonstrate superior performance with a peak signal-to-noise ratio (PSNR) of 35.9377 and intersection over union (IoU) of 0.8677, outperforming conventional methods by 22.5% and deep learning baselines by 11.2%, respectively. Both quantitative and qualitative evaluations demonstrate the robustness and effectiveness of the proposed network in reconstructing 3-D representations of complex anomalies, such as cavities and gangue inclusions, within coal seams. To the best of our knowledge, this is the first deep learning framework that achieves simultaneous resolution enhancement, noise suppression, and 3-D inversion within a unified architecture for coal seam sensing and modeling.