Update on applying machine learning to laboratory shear experimental data

We are applying machine learning to acoustical signals recorded in shear ex- periments conducted on a biaxial shear device located at the Pennsylvania State University (laboratory of Chris Marone). In this work, we show that we can predict frictional failure times (‘labquakes’) in the shear experiments, for both stick-slip and slow slip. This advance is made possible by applying Ran- dom Forest (RF) machine learning to the continuous time series recorded by a single accelerometer listening to the experiment. The RF is trained applying a number of statistical data features over a time interval over which a number of labquakes occur. Remarkably, during testing we find that the RF predicts upcoming failure time at any time in the stress cycle, based only on a short time window of data—a ‘now ’ prediction. The predicted time improves as failure is approached, as other data features add to prediction. In the process of the prediction exercise, we discovered a signal we were previously una- ware of. This signal and its characteristics provide the prediction. The ap- proach should be portable to other failure problems.