11:15 AM - 11:30 AM
[SCG63-03] New Approach Based on Machine Learning for Earthquake Early Warning: Simultaneous Prediction of Noise Discrimination and Threshold Exceedance
Keywords:earthquake early warning, machine learning, recurrent neural network, on-site warning
Traditional earthquake early warning systems process seismometer data through a sequential flow involving noise discrimination, P-wave onset determination, earthquake parameter estimation, ground motion prediction equations, and magnitude updates. This workflow relies on multiple processing steps to issue alerts, leading to significant computational load during earthquakes. Recent advancements in machine learning (ML) have shown potential to improve these systems. For example, Noda and Iwata (2024) applied LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) models for noise discrimination, achieving higher accuracy than traditional methods. Similarly, Noda et al. (2023) and Noda (2024) used convolutional neural networks (CNNs) to enhance earthquake parameter estimation. However, these approaches replace individual steps in the traditional flow and generally increase computational demands, requiring careful implementation to avoid delays.
This study leverages the flexibility of ML to simplify the workflow by integrating noise discrimination and threshold exceedance classification into a single process. Unlike previous methods that categorized data as "seismic" or "non-seismic," our approach focuses on whether the observed ground motion exceeds a predefined threshold. Seismic signals below the threshold are treated as non-seismic (e.g., train vibrations or ambient noise). This allows the model to directly output the probability of threshold exceedance without intermediate steps like parameter estimation or ground motion prediction, significantly reducing computational complexity and enabling consistent processing irrespective of seismic motion.
We developed and tested ML models (LSTM and GRU) using three-component acceleration data. The training data included noise signals (e.g., train vibrations) and seismic waveforms from K-NET and Shinkansen seismometers. The threshold was set at 40 gal for JR-PGA (horizontal two-component acceleration with a 5 Hz low-pass filter; Nakamura et al., 2005). Labels were assigned as 0 for no exceedance and 1 for exceedance. Sequences of 1-second acceleration data were input to the models, which output classification probabilities. Because no significant difference in accuracy was observed between LSTM and GRU, results from the LSTM model are presented here. Classifications were based on a probability threshold of 0.5.
The model demonstrated high precision in distinguishing non-seismic data, achieving a recall rate of 99.9996%. This reflects the model's ability to accurately exclude noise, as JR-PGA values for noise data rarely exceed 40 gal. For seismic signals below the threshold, the recall rate was 88.6%. Misclassifications in this category were primarily observed for records with JR-PGA values near the threshold (average values: 5.3 gal for correct classifications and 24.2 gal for misclassifications). For seismic signals exceeding the threshold, the recall rate was 99.6%. Notably, 54% of threshold exceedance events were identified within 3 seconds of the P-wave onset, compared to only 4.3% based on traditional JR-PGA calculations. This suggests that the LSTM model predicts future amplitudes from P-wave characteristics, allowing earlier warning issuance than conventional methods.
The proposed approach corresponds to on-site warnings, where alerts are generated based on local observations. Future work will focus on extending this method to regional warnings, covering broader areas with enhanced prediction capabilities.
This study leverages the flexibility of ML to simplify the workflow by integrating noise discrimination and threshold exceedance classification into a single process. Unlike previous methods that categorized data as "seismic" or "non-seismic," our approach focuses on whether the observed ground motion exceeds a predefined threshold. Seismic signals below the threshold are treated as non-seismic (e.g., train vibrations or ambient noise). This allows the model to directly output the probability of threshold exceedance without intermediate steps like parameter estimation or ground motion prediction, significantly reducing computational complexity and enabling consistent processing irrespective of seismic motion.
We developed and tested ML models (LSTM and GRU) using three-component acceleration data. The training data included noise signals (e.g., train vibrations) and seismic waveforms from K-NET and Shinkansen seismometers. The threshold was set at 40 gal for JR-PGA (horizontal two-component acceleration with a 5 Hz low-pass filter; Nakamura et al., 2005). Labels were assigned as 0 for no exceedance and 1 for exceedance. Sequences of 1-second acceleration data were input to the models, which output classification probabilities. Because no significant difference in accuracy was observed between LSTM and GRU, results from the LSTM model are presented here. Classifications were based on a probability threshold of 0.5.
The model demonstrated high precision in distinguishing non-seismic data, achieving a recall rate of 99.9996%. This reflects the model's ability to accurately exclude noise, as JR-PGA values for noise data rarely exceed 40 gal. For seismic signals below the threshold, the recall rate was 88.6%. Misclassifications in this category were primarily observed for records with JR-PGA values near the threshold (average values: 5.3 gal for correct classifications and 24.2 gal for misclassifications). For seismic signals exceeding the threshold, the recall rate was 99.6%. Notably, 54% of threshold exceedance events were identified within 3 seconds of the P-wave onset, compared to only 4.3% based on traditional JR-PGA calculations. This suggests that the LSTM model predicts future amplitudes from P-wave characteristics, allowing earlier warning issuance than conventional methods.
The proposed approach corresponds to on-site warnings, where alerts are generated based on local observations. Future work will focus on extending this method to regional warnings, covering broader areas with enhanced prediction capabilities.