A Deep-Learning P-Wave Arrival Picker for Laboratory Acoustic Emissions: Model Training and Its Performance

The acoustic emission (AE) technique has been widely used in studying the cracking and frictional behavior of rocks in laboratory rock mechanics tests. Picking P-wave arrivals of AE waveforms is pivotal in analyzing AE data at an advanced level such as event localization, conducting acoustic tomography, and solving focal mechanisms. Deep-learning-based P-wave arrival pickers have outperformed traditional non-machine-learning algorithms in seismology, enhancing efficiency over manual picking. However, determining the optimal training data set size for these models in laboratory settings has been lacking. This evaluation is crucial for improving AE data analysis efficiency and ensuring the availability of open AE data sets with manually picked waveforms, currently limited by manual picking time constraints. We compiled ~ 50,000 manually picked AE waveforms from a triaxial compression test on a Berea sandstone to create a data set. Introducing AE-PNet, a deep-learning P-wave arrival picker based on modified PhaseNet architecture, we investigated the minimum number of manually picked waveforms needed for effective AE-PNet training. Our study revealed that a minimum of 1500 manually picked waveforms is necessary for adequate AE-PNet generalization, irrespective of total AE waveforms in data sets that AE-PNet processes. AE-PNet consistently outperformed the Akaike Information Criterion picker, even in picking low signal-to-noise waveforms.