A BOA-LGBM framework driven by acoustic emission Mel-frequency cepstral coefficients for early identification of coal-rock fracture processes

Dynamic geotechnical hazards, such as rockbursts and coal–gas outbursts, threaten underground engineering safety. Accurate identification of coal–rock fracture processes is therefore essential for hazard monitoring and early warning. This study proposes an acoustic-emission-based identification framework integrating Mel-frequency cepstral coefficients (MFCCs) with time-domain features. Acoustic emission signals generated during uniaxial compression of coal and sandstone specimens were collected, and 1st–12th-order MFCCs and representative time-domain parameters were extracted using a unified windowing strategy. Correlation analysis was conducted to evaluate feature relationships. A Bayesian-optimization-enhanced Attention-MLP-LightGBM model was developed for coal–rock fracture classification, while SHAP analysis was used to quantify feature importance and guide dimensionality reduction. The proposed model achieved a classification accuracy above 98%. After SHAP-based feature selection, the training time was substantially reduced with negligible performance loss. Independent validation further demonstrated satisfactory robustness and generalization ability. The proposed framework provides an accurate, efficient, and interpretable approach for identifying coal–rock fracture processes and offers potential support for dynamic hazard monitoring and early warning in underground mining.