Investigation on Prediction of Sandstone Failure Under Uniaxial Compression Based on Supervised Deep Learning

Deep learning technique is an effective method for representation of complex relationships applied in various subjects and areas. Acoustic emission technique, a powerful non-destructive testing method, has been used widely to collect acoustic emission signals and monitor rock deformation and damage. This study proposed a deep learning approach to model the relationship between the remaining time to failure and data monitored by acoustic emission technique and predict failure time of sandstone in the uniaxial compression loading experiment. In the meantime, optimal feature combination of model input and optimal parameters of the model structure were obtained and discussed. Three different types of models were established, including fully connected neural networks, gated recurrent units*** and mixed model combining convolutional neural networks with gated recurrent units. The mixed model's coefficient of determination for sandstone failure prediction under uniaxial compression was 0.9699 and its mean absolute percentage error 11.64%. Its performance surpasses those of the others on tested samples. Furthermore, the metric coefficient of determination exhibits significant improvement compared with previous studies. Therefore, the obtained metric results demonstrate a satisfactory level of accuracy for predicting sandstone failure in laboratory-scale applications. Finally, this work demonstrates the capability of deep learning to predict sandstone failure via acoustic emission monitoring technique and offers the theoretical reference for practical applications at the engineering scale.