Auditory Perception-Integrated Depthwise Separable Convolution-Transformer Hybrid Network for Coal-Gangue Ratio Recognition

Abstract In fully mechanized top-coal caving mining, precise recognition of coal samples with varying gangue ratios is a critical link for boosting top-coal utilization efficiency. Yet underground environments are plagued by widespread noise interference, and samples with different gangue ratios often exhibit ambiguous, indistinct features—key challenges hindering accurate identification. This paper proposes a recognition method integrating human auditory perception and a hybrid neural network architecture: First, a auditory perception-like enhancement model tailored to coal and gangue characteristics is built, which enhances feature differences between samples via frequency response regulation and category-specific adjustments. On this basis, a depthwise separable convolution-transformer hybrid network is designed to synergistically learn local subtle features and global dependency relationships. Additionally, a dynamic hard example mining strategy is introduced to strengthen discrimination of ambiguous samples, while a multi-level anti-noise mechanism boosts robustness against intense industrial noise. Experiments confirm effectiveness: at -10 dB SNR, in 5-category recognition, macro-average F1-score reaches 93.68% and overall accuracy 88.10%, verifying the method’s outstanding capability.