Deep learning-based multi-parameter early warning model under true triaxial conditions

There are still many challenges to achieve accurate early warning of rock instability in underground excavation projects. The cusp mutation model combined with energy time series shows great potential in early warning, but this approach has two drawbacks: On the one hand, energy definition method violates the basic thermodynamic law. On the other hand, there is still a large subjectivity in the selection of the energy mutation time series. This results in a large error in the defined warning interval. In this paper, we use machine deep learning method instead of manual single selection of feature values, which greatly avoids the subjective influence of sudden change time series. Based on dynamic moving window method with improved convolutional neural network (CNN) method, a multi-parameter time series prediction of acoustic emission was established. The input features of the three convolutional layers verify and learn each other to strengthen the predictive properties of time series; The dynamic moving window method shortened the time step, deepened the visualization of mutation in the time series, finally the model realized the integration of prediction and warning. The training results show that the multi-parameter warning model has great accuracy, and the selection of warning points is highly consistent with the positioning information; The warning duration stabilizes within a reasonable range as the number of training sessions rises, and the brittleness degree of rock mass has a controllable influence on the duration. This study successfully constructed an integrated prediction and warning model for rock mass instability, which passed the test and performed well on the whole. The model is proposed to provide a more cutting-edge and improved idea for early warning of subsurface rock instability damage.