Abstract: The study of rock stability state judgment and failure time prediction is of great significance for the prevention and control of coal mine dynamic disasters. To study the method of identifying the stable state of rocks and predicting the failure time, a CNN-LSTM Attention convolutional model was constructed by integrating the displacement coordination coefficient index and axial stress curve. The recognition effect of different input steps on the stable state of rocks under this convolutional model was compared. Furthermore, the contribution of each component of the convolutional model was evaluated through ablation experiments. Finally, a rock failure time prediction method based on the displacement coordination coefficient index and axial stress curve was proposed. The results show that: the convolutional model of CNN-LSTM Attention, which integrates displacement coordination coefficient index and axial stress curve, performs the best in rock stability identification task with a recall rate greater than 82.28% when the input step size is 100; The 14 step input has an advantage in training efficiency. The ablation experiment shows that the LSTM module of the model mainly enhances the model's ability to capture features of high-risk areas, the CNN module improves the recognition effect of medium risk areas, and the Attention mechanism can synergistically optimize accuracy and recall, achieving a comprehensive and balanced improvement of the model's ability to identify rock stability. For different rock the convolutional model can predict the final failure time of the rock within 20 seconds after triggering, with an error rate of less than 9%. The research results provide a new and effective method for the identification of rock stability and the prediction of failure time, and can provide a scientific idea for early warning and risk assessment of coal mine dynamic disasters.