On-site Earthquake Early Warning Research Articles

Ground-Shaking Intensity Prediction for Onsite Earthquake Early Warning Using Deep Learning

Abstract Damage caused by an earthquake is highly correlated with the shaking intensity in an area. Therefore, it is beneficial to know the shaking intensity of a certain area in advance so that proper preparations and responses can be made. This article proposes a deep learning approach to predict the onsite intensity of ground shaking for a certain location at the very beginning of an earthquake, namely ground-shaking intensity prediction (GSIP). GSIP learns the features hidden in the acceleration waveform and frequency spectrum during a short initial window after the P-wave arrival. It then predicts the intensity level of the ground shaking based on the extracted features. Traditional methods determine parameters or thresholds by experience, but it is very difficult to select appropriate thresholds, and this may need careful calibration. In contrast, GSIP determines the appropriate parameters from the waveform to predict intensity levels without the need for calibration. The waveforms used for model training and validation were collected from 1991 to 2020 data from Taiwan, 2004 to 2020 data from Japan, and 2005 to 2020 data from Italy. The results show that GSIP can achieve more than 85% accuracy at predicting the intensity level, with a tolerance of one level of intensity, as well as high accuracy and recall. Recent events that occurred in Taiwan in 2021 and 2022 were used to evaluate GSIP, and the results confirm its ability to accurately predict intensity levels in different areas. In addition, GSIP is undergoing testing on the earthquake monitoring system in the Central Weather Bureau of Taiwan and has effectively provided real-time early warning for earthquakes.

On-Site Earthquake Early Warning Model for Selected Records in the NGA-West2 Dataset Using S- and P-Wave Spectral Ratios

ABSTRACT On-site earthquake early warning (EEW) requires the best estimates of earthquake magnitude and distance parameters within a few seconds after the P-wave arrival for estimating the subsequent S-wave parameters. The errors in the estimates of the earthquake P-wave parameters will propagate into the estimates of the S-wave parameters. To solve this problem, we used the methodology by Zhao and Zhao (2019), which uses the spectral ratio R3P between the response spectral values of the first 3 s of the S wave and that of the first 3 s of the P wave, referred to as the R3P model. The modeling presented here was based on strong-motion records from the Next Generation Attenuation (NGA)-West2 dataset. We also used the spectral ratio RFP between the response spectral values from the full records from the S-wave arrival time to the end of the record and that of the first 3 s P wave to develop a second EEW model (the RFP model). An advantage of these two models is that the magnitude and hypocentral distance are not required in considerable magnitude and distance ranges. This means that the errors in the estimated source and path parameters from the first 3 s of the P wave will not affect the model predictions. A theoretical justification for these results is that the magnitude and the distance scaling rates for the first 3 s of the P wave are similar to those of the first 3 s of the S wave. This may also apply to the full S-wave window within useful EEW magnitude and distance ranges. We also found that when the estimated magnitude and distance for a record are necessary, the effect of the corresponding errors would be smaller than using a ground-motion prediction equation (GMPE), because the magnitude and distance scaling rates from this study are smaller than those of many GMPEs.

Interpretability and spatial efficacy of a deep-learning-based on-site early warning framework using explainable artificial intelligence and geographically weighted random forests

Earthquakes pose significant risks globally, necessitating effective seismic risk mitigation strategies like earthquake early warning (EEW) systems. However, developing and optimizing such systems requires thoroughly understanding their internal procedures and coverage limitations. This study examines a deep-learning-based on-site EEW framework known as ROSERS (Real-time On-Site Estimation of Response Spectra) proposed by the authors, which constructs response spectra from early recorded ground motion waveforms at a target site. This study has three primary goals: (1) evaluating the effectiveness and applicability of ROSERS to subduction seismic sources; (2) providing a detailed interpretation of the trained deep neural network (DNN) and surrogate latent variables (LVs) implemented in ROSERS; and (3) analyzing the spatial efficacy of the framework to assess the coverage area of on-site EEW stations. ROSERS is retrained and tested on a dataset of around 11,000 unprocessed Japanese subduction ground motions. Goodness-of-fit testing shows that the ROSERS framework achieves good performance on this database, especially given the peculiarities of the subduction seismic environment. The trained DNN and LVs are then interpreted using game theory-based Shapley additive explanations to establish cause-effect relationships. Finally, the study explores the coverage area of ROSERS by training a novel spatial regression model that estimates the LVs using geographically weighted random forest and determining the radius of similarity. The results indicate that on-site predictions can be considered reliable within a 2–9 km radius, varying based on the magnitude and distance from the earthquake source. This information can assist end-users in strategically placing sensors, minimizing blind spots, and reducing errors from regional extrapolation.

Peak ground acceleration prediction for on-site earthquake early warning with deep learning

Rapid and accurate prediction of peak ground acceleration (PGA) is an important basis for determining seismic damage through on-site earthquake early warning (EEW). The current on-site EEW uses the feature parameters of the first arrival P-wave to predict PGA, but the selection of these feature parameters is limited by human experience, which limits the accuracy and timeliness of predicting peak ground acceleration (PGA). Therefore, an end-to-end deep learning model is proposed for predicting PGA (DLPGA) based on convolutional neural networks (CNNs). In DLPGA, the vertical initial arrival 3–6 s seismic wave from a single station is used as input, and PGA is used as output. Features are automatically extracted through a multilayer CNN to achieve rapid PGA prediction. The DLPGA is trained, verified, and tested using Japanese seismic records. It is shown that compared to the widely used peak displacement (Pd) method, the correlation coefficient of DLPGA for predicting PGA has increased by 12–23%, the standard deviation of error has decreased by 22–25%, and the error mean has decreased by 6.92–19.66% with the initial 3–6 s seismic waves. In particular, the accuracy of DLPGA for predicting PGA with the initial 3 s seismic wave is better than that of Pd for predicting PGA with the initial 6 s seismic wave. In addition, using the generalization test of Chilean seismic records, it is found that DLPGA has better generalization ability than Pd, and the accuracy of distinguishing ground motion destructiveness is improved by 35–150%. These results confirm that DLPGA has significant accuracy and timeliness advantages over artificially defined feature parameters in predicting PGA, which can greatly improve the effect of on-site EEW in judging the destructiveness of ground motion.

Application of the on-site P-wave earthquake early warning method based on site-specific ratios of S-waves to P-waves to the 2016 Kumamoto earthquake sequence, Japan

The on-site P-wave earthquake early warning (EEW) based on the site-specific ratios of S-waves to P-waves has been applied to large-sized offshore earthquakes, and the efficiency of the method has been validated. However, the method requires the P-waves including earthquake ground motions radiated from a large slip area while avoiding the inclusion of S-waves. In this study, we investigated the applicability of the on-site P-wave EEW method for ground motions near an earthquake source fault region, using strong-motion data observed during the 2016 Kumamoto earthquake sequence in Japan. At first, we examined the appropriate time-window length following the arrival of the P-waves. As a result, P-waves with a time-window length of 2.56 s after the arrival at most strong-motion stations were required at least to predict appropriately S-waves for the 2016 Kumamoto earthquake sequence, including the large-sized earthquakes. On the other hand, in the case of the large-sized earthquake as the mainshock (Mj 7.3), the method can predict within a brief time of 0.5 to 2 s in the operational use that strong ground motions exceeding a certain threshold (e.g., acceleration of 150 cm/s2) will come. Moreover, we found that the method was not strongly affected by the non-linearity of soil deposits due to strong ground motions during the 2016 Kumamoto earthquake sequence. The variability of the relationship between P- and S-waves at the seismic bedrock influenced by the source and path effects is larger than the variability of the relationships between P-/S-waves at the seismic bedrock and at the ground surface by the site effects, and therefore, it hides the effect of the non-linearity of soil deposits.Graphical

P-Alert earthquake early warning system: case study of the 2022 Chishang earthquake at Taitung, Taiwan

A series of earthquakes that struck Taiwan's southern Longitudinal Valley on September 17 and 18, 2022 severely damaged several buildings in Taitung and Hualien. The Chishang earthquake, which had a magnitude of ML 6.8 and a large foreshock with a magnitude of ML 6.6 the day before, was the mainshock in this sequence. The strongest intensity reported in the epicentral region during this earthquake sequence, which was 6 + , is the highest ever recorded since the Central Weather Bureau (CWB, renamed as the Central Weather Administration since September 15, 2023) revised its seismic intensity scale. National Taiwan University (NTU) has operated a low-cost earthquake early warning (EEW) system known as the P-Alert for a decade. In this study, we demonstrate the performance of the P-Alert network during the 2022 Chishang earthquake and the largest foreshock. The P-Alert network plotted shake maps during these earthquakes that displayed various values within 5 min. The high shaking areas on these maps were in good agreement with observed damages during this earthquake, providing valuable insights into rupture directivity, a crucial component of earthquake engineering. Individual P-Alert stations acted as on-site EEW systems and provided a lead time of 3–10 s within the blind zone of CWB. For the ML 6.8 mainshock, there was a lead time of at least 5 s, even up to 10 s, demonstrating their effectiveness in the blind zone. The P-Alert regional EEW system provided the first report about 9 s and 7 s after the mainshock and the largest foreshock occurrence, respectively, with estimated magnitudes of 5.74 and 5.67. The CWB system estimated magnitudes of 6.72 and 6.16 in the first report, respectively, about 7 s and 9 s after the earthquake occurrence. The timeliness of the two systems were not significantly different. Despite the effectiveness of the P-Alert network, data loss due to connection interruptions prompted us to develop a new compact data logger for improved data availability.

Real-time prediction of distance and PGA from P-wave features using Gradient Boosting Regressor for on-site earthquake early warning applications

SUMMARY On-site earthquake early warning (EEW) systems represent an important way to reduce seismic hazard. Since these systems are fast in providing an alert and reliable in the prediction of the ground motion intensity at targets, they are particularly suitable in the areas where the seismogenic zones are close to cities and infrastructures, such as Central Italy. In this work, we use Gradient Boosting Regressor (GBR) to predict peak ground acceleration (PGA), and hypocentral distance (D) starting from P-wave features. We use two data sets of waveforms from two seismic sequences in Central Italy: L'Aquila sequence (2009) and the Amatrice–Norcia–Visso sequence (2016–2017), for a total of about 80 000 three-component waveforms. We compute 60 different features related to the physics of the earthquake using three different time windows (1 s, 2 s and 3 s). We validate and train our models using the 2016–2017 data sets (the bigger one) and we test it on the 2009 data set. We study the performances of GBR predicting D and PGA in terms of prediction scores, finding that the models can well predict both targets even using 1 s window, and that, as expected, the results improve using longer time windows. Moreover, we perform a residual analysis on the test set finding that the PGA can be predicted without any bias, while the D prediction presents a correlation with the moment magnitude. In the end, we propose a prototype for a probabilistic on-site EEW system based on the prediction of D and PGA. The proposed system is a threshold-based approach and it releases an alert on four possible levels, from 0 (far and small event) to 3 (close and strong event). The system computes the probability related to each alert level. We test two different set of thresholds: the Felt Alert and the Damage Alert. Furthermore, we consider the lead time (LT) of the PGA to distinguish between useful alerts (positive LT) and Missed Alerts (MA). In the end, we analyse the performance of such a system considering four possible scenarios: Successful Alert (SA), Missed Alert (MA), Overestimated Alert (OA) and Underestimated Alert (UA). We find that the system obtains SA rate about 80 per cent at 1 s, and that it decreases to about 65 per cent due to the increase in MA. This result shows how the proposed system is already reliable at 1 s, which would be a huge advantage for seismic prone regions as Central Italy, an area characterized by moderate-to-large earthquakes (Mw < 7).

Hierarchical Bayesian calibration of regional moment magnitude predictive models

Onsite Earthquake Early Warning (EEW) applications require robust predictive models that can relate the early part of the recorded seismic waves to seismicity and/or ground shaking characteristics, the latter representing the variables used to make decisions about raising or not an alarm. The use of limited information to establish these models within a real-time estimation setting, typically corresponding to the first few seconds after the identification of the seismic wave arrival, introduces significant uncertainty (variability) in the predictions. To increase the number of historical observations available for the model calibration, data from multiple seismic stations within the same geographic region are utilized for the latter task. Unfortunately, this introduces additional sources of variability in the predictive model development, originating from local geographical effects and site conditions that affect differently the underlying physics at each station. This paper introduces a hierarchical Bayesian model updating framework to address this challenge. To better focus the research effort, the calibration of earthquake magnitude nonlinear regression models based on the seismic maximum predominant period is investigated, though underlying principles can be extended to other type of decision variables and intensity measures used in EEW applications. Within the hierarchical updating framework, the predictive models for each station are separately calibrated (first level), while sharing information across the stations by selecting common calibration hyper-parameters (second level). This facilitates a reduction of the variability of the estimates originating from regional wave propagation effects while simultaneously providing sufficient amount of data for calibrating the predictive models even for stations with scarce available records. To support computational efficiency in the hierarchical calibration, conjugate priors are utilized, while for accommodating nonlinear regression relationships samples from the posterior distribution, representing the model calibration, are obtained using Metropolis-within-Gibbs sampling. As illustrative case study, the application to the Sichuan region of Southwestern China is examined, with extensive comparisons performed for the accuracy and robustness of the resultant probabilistic predictive models.

Stable operation process of earthquake early warning system based on machine learning: trial test and management perspective

Earthquake Early Warning (EEW) is an alert system, based on seismic wave propagation theory, to reduce human casualties. EEW systems mainly utilize technologies through both network-based and on-site methods. The network-based method estimates the hypocenter and magnitude of an earthquake using data from multiple seismic stations, while the on-site method predicts the intensity measures from a single seismic station. Therefore, the on-site method reduces the lead time compared to the network-based method but is less accurate. To increase the accuracy of on-site EEW, our system was designed with a hybrid method, which included machine learning algorithms. At this time, machine learning was used to increase the accuracy of the initial P-wave identification rate. Additionally, a new approach using a nearby seismic station, called the 1+ α method, was proposed to reduce false alarms. In this study, an on-site EEW trial operation was performed to evaluate its performance. The warning cases for small and large events were reviewed and the possibility of stable alert decisions was confirmed.

On-Site Earthquake Early Warning System as an Alternative Earthquake Mitigation Solution in Indonesia

The Banten earthquake, which had a magnitude of 6.6 on January 14, 2022, damaged 3,078 houses. That<br />number consisted of 395 heavily damaged units, 692 moderately damaged units and 1,991 lightly damaged units<br />(Tempo.co, 2022). The Banten earthquake was a strong earthquake where the magnitude was greater than a scale of<br />5. Damage to houses caused by the earthquake occurred in most single-story houses or low-rise buildings. Given the<br />large number of one-story houses that are damaged every time a major earthquake occurs in Indonesia, there needs to<br />be appropriate mitigation measures to reduce the risk of earthquake disasters, especially for human casualties. An<br />On-site Earthquake Early Warning System (On-site EEWS) can be an alternative in reducing victims of the disaster.<br />This earthquake early warning system has sensors that are installed on the site of building houses and can predict<br />strong earthquake waves that are destructive in nature (S/Secondary Waves) through P/Primary Waves that arrive<br />early in about 10-20 seconds. This time is sufficient for evacuation for the occupants of a one-story house if the early<br />warning alarm is properly responded to. This early warning radius can reach 20 km from the on-site EEWS location<br />considering that this area has relatively the same vibration effect. Currently, Indonesia through the BMKG is<br />developing EEWS as a part of the existing earthquake mitigation system. The purpose of this study is to describe the<br />application of an on-site earthquake early warning system as an alternative solution for earthquake mitigation in<br />Indonesia. This study evaluates several EEWS applications in the literature to find the best alternative to be applied<br />in Indonesia. The critical factors for on-site implementation of the EEWS discussed in this paper are compared with<br />the Taiwan regional EEWS. Based on the existing validation, the on-site EEWS has an 80% accuracy rate in<br />predicting the intensity level of a strong earthquake, capable to automatically send an alarm message within 3<br />seconds and providing a warning time of at least 8 seconds before a destructive peak S wave arrives.

Unveiling the potential: preliminary result on an AI-based onsite earthquake early warning system deployment in Denpasar City

Due to Indonesia’s geographical location within the Pacific Ring of Fire, it is susceptible to earthquake occurrences. Therefore, the implementation of an Earthquake Early Warning (EEW) System is crucial to mitigate the impact of earthquakes. This study focused on deploying an on-site EEW system on Bali Island, an Indonesian region known for its elevated seismic activity. The study evaluated the system’s performance over five months in 2023. Throughout this observation period, the AI-based on-site EEW system successfully detected the earthquake’s P-wave before the occurrence of stronger seismic waves, such as the peak ground acceleration (PGA). It provided a warning time of up to 75 seconds and accurately predicted the intensity of the impending earthquake. These outcomes meet the fundamental requirements for an effective EEW system. The results obtained from the deployed AI-based on-site EEW system demonstrate the potential for wider implementation of such systems across Indonesia.

Early Peak Ground Acceleration Prediction for On-Site Earthquake Early Warning Using LSTM Neural Network

On-site earthquake early warning techniques, which issue alerts based on seismic waves measured at a single station, are promising, and have performed quite successfully during some damaging earthquakes. Conventionally, most existing techniques extract several P-wave features from the first few seconds of seismic waves after the trigger to predict the intensity or destructiveness of an incoming earthquake. This type of technique neglects the behavior of temporal varying features within P waves. In other words, the characteristics of data sequences are not considered. In this study, a long short-term memory (LSTM) neural network, which is capable of learning order dependence in seismic waves, is employed to predict the peak ground acceleration (PGA) of the coming earthquake. A dense LSTM architecture is proposed and a large data set of earthquakes is used to train the LSTM model. The general performance of the LSTM model indicated that the predicted PGA values are quite promising but are generally overestimated. However, the predicted PGA of the Chi-Chi earthquake data set, whose fault rupture is complex and long, using the proposed LSTM model is more accurate than the PGA predicted in a previous study using a support vector regression approach. In addition, an alternative alert criterion, which issues alerts when the predicted PGA exceeds the threshold in successive time windows, is presented, and the performance of the proposed LSTM model when different PGA thresholds are considered is also discussed.

Seismic Intensity Estimation Using Multilayer Perceptron for Onsite Earthquake Early Warning

Automated early preventive measures are necessary to reduce secondary hazards in case of high-intensity earthquakes, as these have the potential to damage civil structures over a larger distance from the epicentre. The warning in case of a low-intensity earthquake can, however, disrupt human life and cause unnecessary losses. Hence, in this article, a Multilayer Perceptron-classifier is modeled to provide ensuing severity-based warning by predicting the possibility of the onsite intensity exceeding a pre-trained Peak Ground Acceleration threshold related to damaging intensities in the MMI scale. The supervised learning of the classifier utilized seismic features extracted from the strong-motion signal after the onset of the p-wave. A stratified differential feature-window resampling is applied, which enhanced the F-score of intensity class prediction from 79% to 91%. The trained model is realized on a Faster Than Real-Time (FTRT) simulation environment and tested with sliding windows for estimation of earthquake intensity classes for early warning, predicting with 89% accuracy. A seismic dataset consisting of around eight thousand non-earthquake seismic events from the National Capital Region, India, is generated using a fleet of developed seismic sensing nodes (SSNs) over a period of four years. The cross-dataset validation on this generated dataset showed 95.98% accuracy. A novel Factor of Early Warning (FoEW) metric is introduced to measure the effectiveness of warning by normalizing the achieved lead time with the corresponding maximum warning time. Mean FoEW exceeded 85% for the earthquakes with moderate to heavy damage potentials, i.e., MMI ≥ VI, suitable for onsite earthquake early warning.

Neural Network-Based Strong Motion Prediction for On-Site Earthquake Early Warning

Developing on-site earthquake early warning systems has been a challenging problem because of time limitations and the amount of information that can be collected before the warning needs to be issued. A potential solution that could prevent severe disasters is to predict the potential strong motion using the initial P-wave signal and provide warnings before serious ground shaking starts. In practice, the accuracy of prediction is the most critical issue for earthquake early warning systems. Traditional methods use certain criteria, selected through intuition or experience, to make the prediction. However, the criteria thresholds are difficult to select and may significantly affect the prediction accuracy. This paper investigates methods based on artificial intelligence for predicting the greatest earthquake ground motion early, when the P-wave arrives at seismograph stations. A neural network model is built to make the predictions using a small window of the initial P-wave acceleration signal. The model is trained by seismic waves collected from 1991 to 2019 in Taiwan and is evaluated by events in 2020 and 2021. From these evaluations, the proposed scheme significantly outperforms the threshold-based method in terms of its accuracy and average leading time.

Using LSTM Neural Networks for Onsite Earthquake Early Warning

Abstract Onsite earthquake early warning (EEW) systems determine possible destructive S waves solely from initial P waves and issue alarms before heavy shaking begins. Onsite EEW plays a crucial role in filling in the blank of the blind zone near the epicenter, which often suffers the most from disastrous ground shaking. Previous studies suggest that the peak P-wave displacement amplitude (Pd) may serve as a possible indicator of destructive earthquakes. However, the attempt to use a single indicator with fixed thresholds suffers from inevitable errors because the diversity in travel paths and site effects for different stations introduces complex nonlinearities. In addition, the short warning time poses a threat to the validity of EEW. To conquer the aforementioned problems, this study presents a deep learning approach employing long short-term memory (LSTM) neural networks, which can produce a highly nonlinear neural network and derive an alert probability at every time step. The proposed LSTM neural network is then tested with two major earthquake events and one moderate earthquake event that occurred recently in Taiwan, yielding the results of a missed alarm rate of 0% and a false alarm rate of 2.01%. This study demonstrates promising outcomes in both missed alarms and false alarms reduction. Moreover, the proposed model provides an adequate warning time for emergency response.

Applicability of On-Site P-Wave Earthquake Early Warning to Seismic Data Observed During the 2011 Off the Pacific Coast of Tohoku Earthquake, Japan

In this study, the on-site P-wave earthquake early warning (EEW) based on the site-specific spectral ratio of S-wave to P-wave to efficiently incorporate the site characteristics, which can potentially issue the earthquake warning by the time of Ts-p, was developed. The spectral ratio of S-wave to P-wave that are related to the source effects, the path effects, and the site effects are significantly affected by the site effects contrast to the source effects and the path effects in practical. At first, the on-site P-wave EEW method which multiplies a site-specific spectral ratio of S-wave to P-wave prepared in advance by P-wave observed in the real-time at seismic stations is applied to seismic data for moderate-sized earthquakes with a magnitude (Mj) of 5.0–6.0, occurred in the eastern Japan, observed at both the sedimentary basin site and the rock site. As a result, this method predicted well the observed S-wave in the single indicator of SI within the logarithmic standard deviation of 0.25 as well as in the frequency of more than 0.5 Hz. It is, also, confirmed that the site-specific spectral ratio of S-wave to P-wave at a seismic station was stably retrieved from 20 data samples at least. To investigate the applicability of this method to earthquake ground motions induced by a large-scaled earthquake, finally, this method is applied to seismic data during the 2011 off the Pacific coast of Tohoku earthquake, Japan (Mw 9.0). The prediction of S-wave using a time-window of 10 s after P-wave arrived, could not reproduce the observation with the underestimation; however, the prediction of S-wave using a time-window of more than 20 s containing P-wave propagated from an area generating strong motions in the fault, could reproduce the observation. Even in the case of the large-scaled earthquake, the on-site P-wave EEW method based on the site-specific spectral ratio of S-wave to P-wave at a seismic station availably works by using the gradually increasing time-windows after P-wave arrived in the single indicator of SI as well as in the frequency content, avoiding the mixture of S-wave into a part of P-wave.

Earthquake Early Warning System for Structural Drift Prediction Using Machine Learning and Linear Regressors

In this work, we explored the feasibility of predicting the structural drift from the first seconds of P-wave signals for On-site Earthquake Early Warning (EEW) applications. To this purpose, we investigated the performance of both linear least square regression (LSR) and four non-linear machine learning (ML) models: Random Forest, Gradient Boosting, Support Vector Machines and K-Nearest Neighbors. Furthermore, we also explore the applicability of the models calibrated for a region to another one. The LSR and ML models are calibrated and validated using a dataset of ∼6,000 waveforms recorded within 34 Japanese structures with three different type of construction (steel, reinforced concrete, and steel-reinforced concrete), and a smaller one of data recorded at US buildings (69 buildings, 240 waveforms). As EEW information, we considered three P-wave parameters (the peak displacement, Pd, the integral of squared velocity, IV2, and displacement, ID2) using three time-windows (i.e., 1, 2, and 3 s), for a total of nine features to predict the drift ratio as structural response. The Japanese dataset is used to calibrate the LSR and ML models and to study their capability to predict the structural drift. We explored different subsets of the Japanese dataset (i.e., one building, one single type of construction, the entire dataset. We found that the variability of both ground motion and buildings response can affect the drift predictions robustness. In particular, the predictions accuracy worsens with the complexity of the dataset in terms of building and event variability. Our results show that ML techniques perform always better than LSR models, likely due to the complex connections between features and the natural non-linearity of the data. Furthermore, we show that by implementing a residuals analysis, the main sources of drift variability can be identified. Finally, the models trained on the Japanese dataset are applied the US dataset. In our application, we found that the exporting EEW models worsen the prediction variability, but also that by including correction terms as function of the magnitude can strongly mitigate such problem. In other words, our results show that the drift for US buildings can be predicted by minor tweaks to models.

Onsite Early Prediction of PGA Using CNN With Multi-Scale and Multi-Domain P-Waves as Input

Although convolutional neural networks (CNN) have been applied successfully to many fields, the onsite earthquake early warning by CNN remains unexplored. This study aims to predict the peak ground acceleration (PGA) of the incoming seismic waves using CNN, which is achieved by analyzing the first 3 s of P-wave data collected from a single site. Because the amplitude of P-wave data of large and small earthquakes can differ, the multi-scale input of P-wave data is proposed in this study in order to let the CNN observe the input data in different scales. Both the time and frequency domains of the P-wave data are combined into multi-domain input, and therefore the CNN can observe the data from different aspects. Since only the maximum absolute acceleration value of the time history of seismic waves is the target output of the CNN, the absolute value of the P-wave time history data is used instead of the raw value. The proposed arrangement of the input data shows its superiority to the one directly inputting the raw P-wave data into the CNN. Moreover, the predicted PGA accuracy using the proposed CNN approach is higher than the one using the support vector regression approach that employed the extracted P-wave features as its input. The proposed CNN approach also shows that the accuracy of the predicted PGA and the alert performances are acceptable based on data from two independent and damaging earthquakes.

A Framework for Evaluating Earthquake Early Warning for an Infrastructure Network: An Idealized Case Study of a Northern California Rail System

Earthquake early warning (EEW) systems provide a few to tens of seconds of warning before shaking hits a site. Despite the recent rapid developments of EEW systems around the world, the optimal alert response strategy and the practical benefit of using EEW are still open-ended questions, especially in areas where EEW systems are new or have not yet been deployed. Here, we use a case study of a rail system in California’s San Francisco Bay Area to explore potential uses of EEW for rail systems. Rail systems are of particular interest not only because they are important lifeline infrastructure and a common application for EEW around the world, but also because their geographically broad yet networked infrastructure makes them almost uniquely well suited for utilizing EEW. While the most obvious potential benefit of EEW to the railway is to prevent derailments by stopping trains before the arrival of shaking, the lead time for warnings is usually not long enough to significantly reduce a train’s speed. In reality, EEW’s greatest impact is preventing derailment by alerting trains to slow down or stop before they encounter damaged track. We perform cost-benefit analyses of different decision-making strategies for several EEW system designs to find an optimal alerting strategy. On-site EEW provides better outcomes than source-parameter-based EEW when warning at a threshold of 120 gal (the level of shaking at which damage might occur) regardless of false alarm tolerance. A source-parameter-based EEW system with a lower alerting threshold (e.g., 40 gal) can reduce the exposure to potentially damaged track compared to an on-site system alerting at 120 gal, but a lower alerting threshold comes at the cost of additional precautionary system stops. The optimal EEW approach for rail systems depends strongly on the ratio of the cost of stopping the system unnecessarily to the potential loss from traversing damaged tracks.

Deep Learning-Based, Real-Time, False-Pick Filter for an Onsite Earthquake Early Warning (EEW) System

This paper presents a real-time, false-pick filter based on deep learning to reduce false alarms of an onsite Earthquake Early Warning (EEW) system. Most onsite EEW systems use P-wave to predict S-wave. Therefore, it is essential to properly distinguish P-waves from noises or other seismic phases to avoid false alarms. To reduce false-picks causing false alarms, this study made the EEWNet Part 1 'False-Pick Filter' model based on Convolutional Neural Network (CNN). Specifically, it modified the Pick_FP (Lomax et al.) to generate input data such as the amplitude, velocity, and displacement of three components from 2 seconds ahead and 2 seconds after the P-wave arrival following one-second time steps. This model extracts log-mel power spectrum features from this input data, then classifies P-waves and others using these features. The dataset consisted of 3,189,583 samples: 81,394 samples from event data (727 events in the Korean Peninsula, 103 teleseismic events, and 1,734 events in Taiwan) and 3,108,189 samples from continuous data (recorded by seismic stations in South Korea for 27 months from 2018 to 2020). This model was trained with 1,826,357 samples through balancing, then tested on continuous data samples of the year 2019, filtering more than 99% of strong false-picks that could trigger false alarms. This model was developed as a module for USGS Earthworm and is written in C language to operate with minimal computing resources.