Roughness-guided bidirectional Mamba for noise-robust acoustic-vibration coal-gangue recognition under underground equipment noise

Acoustic-vibration coal-gangue recognition is required for intelligent top-coal caving, but non-stationary underground equipment noise can induce distribution shifts between calibration and operating conditions. This paper formulates the task as a noise-scenario generalization problem and proposes RAAF-BiPAC-Mamba, a roughness-guided bidirectional Mamba framework for noise-robust recognition. The framework uses an efficient bidirectional selective state-space backbone to encode 2048-point acoustic-vibration impact responses and introduces two robustness mechanisms. Roughness-aware adaptive fusion (RAAF) uses latent vibration roughness as a modality-reliability cue, whereas the physics-aware constraint with gradient reversal (PAC-GRL) combines gradient-reversal alignment with a label-conditioned, one-sided roughness regularizer. The method is evaluated on a controlled-platform dataset under source-event-disjoint validation, leave-one-noise-scenario-out (LONSO) evaluation, and synthetic held-out equipment-noise overlay testing. Under the primary source-event-disjoint protocol, RAAF-BiPAC-Mamba achieves 93.74% average accuracy across seven noise scenarios and 91.18% under severe mixed noise (S6), improving over Bi-Mamba by +3.25 and +3.75 percentage points and over the adapted HMBCNN-style implementation by +4.00 and +4.20 percentage points, respectively. It also obtains 89.54% Avg-LONSO accuracy and 86.00% accuracy in the held-out mixed-equipment overlay at −6 dB, while using 0.234 M parameters and 0.063 G FLOPs. These results support noise-shift robustness under the evaluated controlled-platform and held-out noise-overlay settings; they do not establish transfer across mine sites, sensor hardware or calibration, geological conditions, or installation configurations.