Data-driven prediction for uniaxial compression failure time of strong bump-prone coal using acoustics-informed neural networks

The failure prediction for coal material is crucial in mining engineering and mining safety. Classical approaches for failure prediction of coal material generally rely on identifying anomalous changes in monitored characteristics as precursors to failure, yet they lack the capability for precise and quantitative forecast. The data science-based deep learning techniques have supplied options to realize the possibility of quantitative and dynamic regression prediction. In this work, data-driven prediction models for remaining time to failure based on acoustics-informed and supervised neural networks were trained from scratch on coal samples under uniaxial compression. Through feature engineering process, feature combination from acoustic emission characteristics was selected. Notably, the RMS (root mean square of signal) was found as a pivotal feature for predicting coal failure. The proposed models on testing dataset were evaluated and compared using various metrics. In accordance to the results, the lightweight and hybrid deep learning model MCFPNet outperformed the other considered models on each performance metric. Among the metrics, the R² of MCFPNet reached 0.9652. Meanwhile, the unique interacting evaluation metric R² also produced performance improvement by 8.45% above the optimal of the other intelligent architectures for rock failure time. Furthermore, the uncertainty analyses via confidence intervals demonstrated the prediction robustness of the MCFPNet in repetitive training and testing process. Therefore, our work substantiates the effectiveness in quantitative prediction using the acoustics-informed and supervised deep learning model. Finally, the proposed coal failure prediction method offers valuable research inspiration and potential for the early warning of mining dynamic disasters.