Many laboratory fault failure experiments are conducted as analogue of earthquakes, in which most of them are coupled with acoustic emission (AE) as a powerful diagnostic tool for investigating failure precursors. The purpose of this study is to predict time to the next failure in a laboratory fault failure experiment before failures happen based on the instantaneous recorded AE signals. A customized deep learning network comprising the convolutional neural network module and the recurrent neural network module is built and trained using raw AE data directly. No statistical characteristics or handmade features are extracted from raw data, avoiding any possible precursor information losses. More than 600 million AE data from a repetitive fault failure experiment are segmented as several thousand equilong sequences to form training and validation samples. The proposed network delivers satisfactory predicted results with the R2 value 0.55, much better than results using traditional earthquake catalogs method. Results of this study also demonstrate that our network does not prioritize those AE signals collected when failures impend, which is a common bias in traditional earthquake prediction methods. This study definitely holds the promise of using deep learning in earthquake prediction. Further studies are needed when analogous studies proceed to an industrial practice.
Read full abstract