A Machine-Learning-Based Automatic P-Wave Detecter and Picker for Acoustic Emission Events in Laboratory Experiments

Abstract:

Deformation experiments conducted under controlled pressure and temperature in laboratory and acoustic emission (AE) events are monitored by transducers during the experiments. Accurate first P-wave arrival times are needed for locating the events. There have been several machine-learning-based methods to pick P-wave arrival times in waveform records of individual stations. Here we developed a new method to detect AE events and pick P-wave arrivals at multiple transducers simultaneously. We treated waveforms recorded by different transducers as a 2D image and applied a convolutional neural network (CNN) to classify the image to detect AE events. We then applied a fully convolutional network (FCN) to do image recognition to pick P-wave arrival times. We tested the method using data from a transformational faulting experiments on Mg2GeO4. P-wave arrival times of 582 AE events were manually picked at 6 transducers. We chose 50 events with P-wave arrival times and randomly cut the waveform data into 100 segments to train our model. The trained model was applied to the waveform data of 532 events. 488 events were detected and picked . 44 events were missed. 96.56% P-wave arrival times were picked by FCN within an error less than 0.05 micro-seconds.