Characterization of the Accuracy by AIC Picking and Manual Picking with the Decaying of Event Magnitude

This is the authors' abstract. We don't add key points for this paper.

Laboratory experimental rock mechanics studies with acoustic emission (AE) recordings are often viewed as analogous to investigations of field-scale phenomena. The application of U-Net for processing field-scale seismic waveform has achieved widely recognized success. Such a technique is tantalizing for experimentalists handling AE signal processing in many ways. However, training a U-Net for processing laboratory AE signal faces many challenges. This study aims to address a few issues encountered on training a U-Net for processing laboratory AE signal. We conduct extensive manual pickings on selected signals of a well-studied laboratory data set. These signals are selected at different intervals along the power-law distribution of AE events. Manual picking consists of two experimental rounds: in the 1st round, two pickers will independently pick the selected signals ranging from the large magnitude events to the small magnitude events. In the 2nd round, two pickers will collaboratively pick these selected signals. Upon these manual picking experiments, we find that the accuracy from an independent manual picker is not reliable. The collaborative manual picking achieves a high matching ratio between the collaborative pickers and the AIC auto-picking. The time consumed in the manual picking experiments also exposes some economic consideration on pursuing a laboratory AE processing U-Net. The condition of high matching ratio in the collaborative manual picking imposes a demand for data augmentation for training the U-Net for laboratory AE signal processing. The discovery of a signal feature confusing the AIC auto-picking process is helpful for designing the data augmentation on the labelled signals for training the laboratory AE signal processing U-Net. 1. INTRODUCTION Laboratory experimental rock mechanics studies with acoustic emission (AE) recording are often viewed as an analogue to investigations of field-scale phenomena, such as natural earthquakes [1-4], or induced seismicity due to hydraulic fracturing or wastewater disposal [5, 6]. Although creating the analogue towards actual field-scale studies is the aim, differences exist in almost every step of the laboratory AE analysis against field-scale seismic data processing. Here we give a few examples along the sequence from data acquisition to signal processing, and to source location. First, the data acquisition sensors used in laboratory settings are mostly recording only one-component waveforms versus the three-component seismometer used in the field. Second, the AIC (Akaike's information criterion) auto-picking is usually applied for laboratory AE signal processing versus the combinative semiauto-manual picking in field-scale seismograms. Third, usually only the P-wave will be used for laboratory source location versus the multiple phases used for field-scale source location. Also, the P-wave velocity is usually predetermined for laboratory source location versus the wave velocity field which can be simultaneously solved in the inverse problem for field-scale studies. Each of these differences lie in series and thusly affect the investigation strategies in the previous and subsequent steps. As such, transferring a well-developed field-scale data analysis technique into laboratory-scale usage must be carefully conducted and be independently verified, and vice versa.