Application of convolutional neural networks for rock stability identification and failure time prediction

The identification of rock stability and the prediction of failure time are crucial for the early warning and prevention of sudden geological disasters such as landslides and collapses. To address these challenges, this study proposes three convolutional prediction models: CNN-LSTM-Attention, CNN-BiLSTM-Attention, and CNN-GRU-Attention. The displacement coordination coefficient (DCC) index and stress curves were employed as input variables to evaluate the performance of each model in discriminating rock stability states under different data structures and input configurations. Furthermore, an innovative methodology for predicting rock failure time utilizing convolutional models was developed. The experimental results demonstrate that the CNN-LSTM-Attention model, utilizing a 10×10×2 data structure, exhibits superior performance in rock stability state discrimination, achieving an accuracy of 95.25% on the validation set and a recall rate of 96% for samples in high-risk areas. Furthermore, when the DCC index was used as the input variable, the CNN-LSTM-Attention model achieved recall rates of 95.8% and 86.5% for medium- and high-risk areas, respectively, in the validation set. These findings indicate that the proposed convolutional models can effectively identify rock stability states by leveraging surface deformation characteristics. The CNN-LSTM-Attention model, with the DCC index as the input variable, is capable of predicting the final rock failure time in real-time once the DCC abrupt change exceeds 0.78. For different rocks, the model can predict the failure time within 20 s after the DCC reaches 0.78, with an error rate of less than 9%. The convolutional neural network model, developed based on the DCC index, provides a novel methodological approach for geohazard early warning research, facilitating slope instability monitoring and earthquake precursor identification using GNSS and other displacement measurement techniques. • A new method for predicting rock failure time using Displacement Coordination Coefficient. • The influence of different data structures and convolutional neural networks on the prediction of rock failure time. • The predictive performance of the optimal model for different rock failure times