11:00 AM - 1:00 PM
[SCG51-P03] Hypocenter Determination Using Deep Learning Technique Based on Arrival Time Probability Trace
Keywords:Deep learning, Phase picking, Earthquake location
The development of seismic event catalogs from continuous waveform data typically involves several processes: event detection, phase picking, event (phase) association, and hypocenter determination. For example, event detection by STA/LTA ratio (Allen 1978), phase picking based on AR-AIC model (Takanami and Kitagawa, 1998), and event association based on grid search approach (Zhang et al. 2019) were used to obtain P and S-wave arrival times for a specific seismic event, and then, their hypocenters were located by, for example, a least squared method. In recent years, machine learning methods, including deep learning, have been applied to each of these processes and have shown high-performance results (e.g., Zhu and Beroza, 2019; Mousavi et al., 2020).
Deep learning is a tool to automatically find feature values from a large amount of training data. If enough training data can be prepared, end-to-end processing will likely yield better results than the case when the procedure is divided into many processes. In the case of earthquake catalog development, it corresponds to developing a network to estimate a hypocenter and an origin time directly from continuous waveforms. In fact, Perol et al. (2018) reported that a hypocenter could be determined from a waveform even of one station. However, in such an approach, hypocenters of events occurring in regions where earthquakes have not been included in the training data set cannot be determined. Although this problem may be solved using synthetic waveforms as training data (Tsuboi and Sugiyama 2019), it is difficult to effectively use high-frequency components that are difficult to model, resulting in the difficulty of automatic processing of a large number of small earthquakes that would benefit greatly from automation.
This study suggests a one-stop analysis method to develop a seismic catalog from continuous waveforms using deep learning technique. The process consists of two deep learning processing: 1) computation of probability traces of arrival times from continuous waveforms, and 2) estimation of hypocenter coordinates and an origin time from the probability traces at many stations. I tested this process by assuming hydraulic fracturing laboratory experiments of Naoi et al. (2020) and Tanaka et al. (2021), where 65 x 65 x 130 mm specimens were used. I used observed waveforms in the actual experiments for the training of the deep learning network of the first process and used probability traces numerically generated from many virtual AE sources for the second process.
For the phase picking process, I adopted an Unet type architecture (Ronneberger et al. 2015), similar to Zhu and Beroza (2019). Waveforms of 1024 samples (10 MHz sampling) around P-wave arrival times were input to the network, and it outputs a trace of probability score of P-wave arrival time. For the training, I used waveforms obtained for AE events whose hypocenters were accurately determined by automatic processing of Naoi et al. (2020) and Tanaka et al. (2021), which is based on STA/LTA and AR-AIC algorithm.
For the hypocenter determination process, we developed another deep learning network that uses the arrival time probability traces of 16 AE sensors as input data and outputs a hypocenter coordinate and an origin time. The arrival time probability traces for the training were synthetically prepared from the virtual hypocenters that were generated based on random numbers. For the trace set, some pulses were deleted, or dummy pulses were added using random numbers. I prepared training data corresponding to 200,000 AE sources, verification data for 20,000 sources, and test data for 10,000 sources for the training. I also made another probability trace dataset that was synthetically generated with arrival time errors obeying a normal distribution with a standard deviation of 0.5 μs for the performance evaluation. I applied the trained network to the synthetic dataset and confirmed that their hypocenters were successfully determined within several millimeter accuracies.
Deep learning is a tool to automatically find feature values from a large amount of training data. If enough training data can be prepared, end-to-end processing will likely yield better results than the case when the procedure is divided into many processes. In the case of earthquake catalog development, it corresponds to developing a network to estimate a hypocenter and an origin time directly from continuous waveforms. In fact, Perol et al. (2018) reported that a hypocenter could be determined from a waveform even of one station. However, in such an approach, hypocenters of events occurring in regions where earthquakes have not been included in the training data set cannot be determined. Although this problem may be solved using synthetic waveforms as training data (Tsuboi and Sugiyama 2019), it is difficult to effectively use high-frequency components that are difficult to model, resulting in the difficulty of automatic processing of a large number of small earthquakes that would benefit greatly from automation.
This study suggests a one-stop analysis method to develop a seismic catalog from continuous waveforms using deep learning technique. The process consists of two deep learning processing: 1) computation of probability traces of arrival times from continuous waveforms, and 2) estimation of hypocenter coordinates and an origin time from the probability traces at many stations. I tested this process by assuming hydraulic fracturing laboratory experiments of Naoi et al. (2020) and Tanaka et al. (2021), where 65 x 65 x 130 mm specimens were used. I used observed waveforms in the actual experiments for the training of the deep learning network of the first process and used probability traces numerically generated from many virtual AE sources for the second process.
For the phase picking process, I adopted an Unet type architecture (Ronneberger et al. 2015), similar to Zhu and Beroza (2019). Waveforms of 1024 samples (10 MHz sampling) around P-wave arrival times were input to the network, and it outputs a trace of probability score of P-wave arrival time. For the training, I used waveforms obtained for AE events whose hypocenters were accurately determined by automatic processing of Naoi et al. (2020) and Tanaka et al. (2021), which is based on STA/LTA and AR-AIC algorithm.
For the hypocenter determination process, we developed another deep learning network that uses the arrival time probability traces of 16 AE sensors as input data and outputs a hypocenter coordinate and an origin time. The arrival time probability traces for the training were synthetically prepared from the virtual hypocenters that were generated based on random numbers. For the trace set, some pulses were deleted, or dummy pulses were added using random numbers. I prepared training data corresponding to 200,000 AE sources, verification data for 20,000 sources, and test data for 10,000 sources for the training. I also made another probability trace dataset that was synthetically generated with arrival time errors obeying a normal distribution with a standard deviation of 0.5 μs for the performance evaluation. I applied the trained network to the synthetic dataset and confirmed that their hypocenters were successfully determined within several millimeter accuracies.