Abstract: The prediction of coal rock instability is a significant topic in mining and safety engineering. Classical methods for predicting the instability of coal materials rely on the identification of abnormal monitoring information, lacking precise quantitative and dynamic prediction capabilities. However, deep learning algorithms provide an effective pathway for quantitative regression prediction of coal destruction. This study constructs a dynamic regression prediction method for the remaining time of coal fracturing destruction based on an infrared radiation information-driven lightweight coupled deep neural network. By training a lightweight coupled intelligent architecture on preprocessed infrared radiation time series data, the study automatically extracts key instability features. Through model validation and feature optimization, it is determined that the optimal combination of input features consists of the mean, maximum, minimum, and median temperatures of the infrared images. After testing and evaluating the proposed model on an independent test set, the R² value reaches 0.9895, a 2.52% improvement over the currently known optimal methods for predicting rock destruction time. The MAPE is 9.78%, a 42.30% improvement over the optimal results evaluated using this indicator. These results validate the efficiency of the infrared feature deep extraction algorithm. The uncertainty analysis of confidence intervals shows that the prediction error confidence intervals are generally close to the time axis, and the error amplitude is less than 25 seconds for most periods, proving that the model has strong robustness in repeated training-test results and good prediction stability. This model can accurately dynamically quantify the remaining time of coal instability, providing important scientific research references and application prospects for early warning of mining dynamic disasters.