Accurate real-time assessment of underground stress fields is critical for mine safety, yet conventional inversion methods struggle to balance precision with computational efficiency. To address these challenges, we propose the Dual-Branch Multi-Scale Convolutional Attention Network (DB-MS-CAN), which employs a decoupling-and-aggregation strategy. The framework integrates a Dual-Branch Fusion Module (BFM) to isolate sparse transient signals from persistent mining noise and a Multi-Scale Dense Feature Aggregation (MDFA) module to reconstruct complex geological structures across varying spatial scales. To ensure rigorous validation, the model was evaluated on both geologically constrained real-world datasets and an independent synthetic benchmark governed by wave propagation physics. Results demonstrate that DB-MS-CAN achieves a structural similarity index (SSIM) of 0.925 and an root mean square error (RMSE) reduction of approximately 40% compared to traditional ray-based tomography. The proposed model achieved significant performance improvements compared to advanced baselines, including CNN-LSTM and Vision Transformers. Notably, the model maintains high fidelity (correlation coefficient > 0.90) under extreme 0 dB SNR conditions and achieves an inference speed of ~7 seconds per event on a single GPU, outperforming iterative solvers by orders of magnitude. This framework provides a robust and efficient solution for dynamic hazard early warning in deep mining environments.