Earthquake early warning (EEW) systems aim to provide timely alerts before strong shaking reaches populated areas. Current EEW systems often rely on multiple modules (data processing, parameter evaluation, alert filtering) and involve complex empirical threshold settings for triggering alarms based on continuous waveform streams. These thresholds present challenges in defining optimal alert criteria, balancing real-time efficiency against false alarms. Single-station methods traditionally use the initial P-waves to estimate parameters, often requiring confirmation from additional stations to reduce false positives. Regional EEW using networks of stations typically necessitates data from at least four triggered stations for accurate parameter determination, adding further delay. Deep learning offers the potential to improve EEW by efficiently extracting information from initial P-waves and handling limited early-stage data. While deep learning-based approaches for earthquake monitoring exist, they often struggle with generalization across different regions and station distributions. Transfer learning is often necessary when applying these models to new areas. This research focuses on developing data augmentation techniques to train generalized neural networks for universal real-time EEW, capable of handling diverse station distributions and regions without the need for complex empirical configurations.

Traditional travel time-based earthquake monitoring involves multiple steps: detection, phase picking, association, location, and magnitude evaluation. Machine learning has improved automation in catalog construction, particularly with neural networks for phase detection and association. However, errors in any step impact final results and increase processing time. EEW demands rapid parameter reporting with limited initial information. Direct waveform data mining using deep learning techniques offers an alternative, bypassing intermediate steps and complex threshold settings. Single-station monitoring using convolutional neural networks (CNNs) has shown promise in earthquake detection and parameter estimation. However, challenges remain in accurately determining back-azimuth due to noise, and network-based methods often require transfer learning for new regions. Previous studies have demonstrated the effectiveness of fully convolutional neural networks and graph neural networks for earthquake location and magnitude estimation, but generalization remains a challenge due to varying station distributions and geological structures. This study aims to address these limitations by creating generalized training earthquakes through data augmentation, resulting in a more robust and universally applicable EEW system.

This study proposes a novel data augmentation method to create generalized earthquakes for training universal neural networks for real-time EEW. Three neural networks are designed: one for detection, one for location, and one for magnitude estimation. The networks are based on fully convolutional architectures. The training data consists of single-station seismograms from earthquakes in Italy, Oklahoma, and Southern California. A key innovation is the recombination of these single-station seismograms to synthesize numerous generalized earthquakes at arbitrary locations within the study area with various station distributions. This recombination process assumes that seismograms with similar epicentral distances and depths exhibit similarity in arrival times, regardless of geographic location. To generate a training sample, the method randomly selects an earthquake location and 4–12 station locations. Seismograms with appropriate epicentral distance and depth are selected from the base dataset for each synthetic station. The E and N components are rotated to match the new azimuth direction using a rotation matrix. The three-component seismograms are filtered (2–8 Hz for detection/location, 0.5–9 Hz for magnitude) and amplitude-normalized using Hutton and Boore’s empirical equation to achieve consistency in magnitude. For detection and location training, 355,001 earthquakes within the monitoring area are generated, plus 2000 outside for anomaly detection. Waveforms are shifted and signal lengths varied in the 30-s time window to simulate real-time scenarios with partially triggered stations. For magnitude estimation, single-station waveforms are used as input. 200,000 samples from the base dataset and 100,000 from a separate set of larger earthquakes are generated via time window truncation and random shifting. The models are trained using the Adam optimizer with a learning rate of 10⁻⁴ and dropout layers. The detection and location networks are merged for simultaneous processing. Geographical coordinates are converted to Cartesian coordinates for processing and then converted back to geographical coordinates for output. The system determines triggered stations based on predicted P-arrival times within the 30-s window. Magnitude estimation uses data from triggered stations, and the final magnitude is the average across these stations.

The trained neural networks were applied to earthquake sequences in Osaka, Japan, and Ridgecrest, California. For real-time simulations, continuous waveforms were fed into the network with a 30-s sliding window and 0.5-s time interval. Earthquake detection and location were determined by exceeding specified probability density function (PDF) thresholds (0.7 and 0.6 respectively). Origin time was estimated by subtracting the travel time from the first P-arrival time. Magnitude was estimated using the mean of triggered stations with high PDFs and signal durations over 2 s. The results demonstrate that the system can reliably report earthquake parameters within 4 seconds of the first P-wave arrival. For the Osaka (M 6.1) and Ridgecrest (M 6.4) main shocks, location errors were 3.9 km and 3.2 km at the 4th second, respectively, and magnitudes were estimated at 5.5 and 5.8 (underestimated). Robustness testing using one day of continuous data from Osaka and Ridgecrest showed the ability to detect and locate numerous events, with low false alarm rates for M > 3.0 earthquakes. Accuracy assessment using selected events revealed epicentral errors of 2.6 km (Osaka) and 4.5 km (Ridgecrest), with magnitude errors of 0.08 and 0.11, respectively. Further analysis showed that early warning could be activated as early as 3–4 s after the first triggered station. Comparison with a traditional picking-based method showed that the neural network approach yielded a 0.7 s reduction in first alarm time and lower location and magnitude errors. Generalization testing across Japan and Northern California, using 139 large earthquakes, resulted in mean errors of 4.9 km (location), 4.0 km (depth), and 0.17 (magnitude) for onshore earthquakes. Offshore earthquakes yielded larger errors (7.3 km, 8.7 km, and 0.22, respectively), potentially due to complex velocity structures. Analysis of first alarms across these regions showed that most could be issued within 4 s for densely monitored events. Sparsely monitored events resulted in longer first alarm times. The mean errors at first alarm were 5.5 km (location), 4.2 km (depth), and 0.32 (magnitude). Finally, application to the M 7.3 Kumamoto earthquake showed initial parameter reporting at 3.5 s with location and magnitude errors of 6.1 km and 1.4, respectively. Comparison with the traditional method showed that the neural network method resulted in a 2-second reduction in the time to the first issuance of the magnitude estimation.

The findings demonstrate the potential for global application of the developed neural networks for real-time EEW. The models' ability to generalize across geographically and temporally diverse regions, with varying station distributions, significantly reduces the need for region-specific training and complex empirical configurations typically required by traditional methods. The consistent performance across different datasets suggests that the networks have learned generalized features of earthquake waveforms, rather than simply memorizing local characteristics. The ability to provide reliable earthquake parameters within a few seconds of the initial P-wave arrival significantly improves the timeliness of EEW systems, offering vital seconds for emergency response and protective actions. The comparison with traditional methods highlights the significant advantages in efficiency and accuracy provided by the proposed deep learning approach.

This study presents a novel approach to real-time earthquake early warning utilizing generalized neural networks trained with recombined seismic data. The resulting models demonstrate robust performance across various regions and station configurations, offering a significant advancement in the field. Future research could explore improvements by incorporating transformer architectures to enhance feature extraction, addressing challenges in distinguishing closely spaced events, and incorporating teleseismic data and smartphone accelerometer data for enhanced robustness and broader coverage. Further investigation into depth estimation, particularly for deep earthquakes, and the inclusion of additional data from various regions and instrument types globally could further improve accuracy and generalizability.

The study's limitations include the assumption of a simplified 1D layered velocity model for data recombination, which might affect accuracy in regions with complex velocity structures. The system may struggle with distinguishing multiple events within the 30-s time window, potentially leading to missed events or false alarms. The reliance on surface stations limits the accuracy of depth estimations, and the performance might be affected by the presence of abnormal signals in the waveforms. Finally, the current method requires that the station distribution covers the monitoring area, which could limit application to offshore monitoring using onshore stations.