基於注意力機制之神經網路運用於現地型地震預警

Abstract

地震預警系統透過爭取在地震發生後的極短時間內提供警報，以減少大地震所引起的傷亡與災損。地震預警分為區域型與現地型，區域型地震預警難以及時對震央附近的區域發出警告，而現地型地震預警可彌補其不足之處，其中P波到達後數秒內的最大位移振幅(Pd)已被視為現地型地震預警的指標，此方法藉由分析Pd是否超過特定閾值以判定是否應預警。然而，由於地震訊號的複雜性高，Pd方法僅以單一閾值預測地表振動將導致不確定性增加，進而使得 Pd方法成為非線性問題。由於機器學習擅於解決非線性問題，前人研究提出以長短期記憶神經網路(Long Short-Term Memory, LSTM)分析P波到達後短時間窗的訊號，並以最大地動加速度達80 gal作為發布警報的依據，再藉由神經網路模型輸出警報機率，決定是否應發布警報，其研究成果有效提升現地型地震預警的表現。然前人研究中指出最大地動加速度易出現異常高值而使不穩定性較高，若以其作為預警的標準，將難以反映實際災害情形，而最大地動速度則與災害相關程度更高，更適合作為災害評估的指標。 而近年來因為注意力機制(Attention mechanism)具有捕捉長期依賴關係、平行計算等特點而使得許多基於注意力機制的神經網路被廣泛應用於自然語言處理任務中，並且在許多任務上取得較LSTM更為卓越的表現。本研究將以注意力機制為基礎建構可運用於現地型地震預警的模型，使用短時間窗的加速度與速度訊號，以及加速度向之測站修正作為輸入資料，並藉由模型提供後續最大地動速度達15 cm/s的機率，以決定是否需發布警報。同時選取臺灣近期四個致災性地震測試模型表現，並以LSTM模型與Pd方法為基準進行比較。注意力機制建構之預警模型整體結果顯示在震央距70公里以內的區域的誤報率為30.3%、漏報率為7.4%，F1 score達到0.694，且提供了平均6.35秒的預警時間，與LSTM與Pd相比，注意力機制的整體表現有所提升，同時亦提供了足夠的預警時間。

The earthquake early warning system aims to provide alerts within an extremely short time after an earthquake occurs, in order to reduce casualties and damage caused by major earthquakes. The earthquake early warning system consists of two types, regional and onsite. Regional earthquake early warning is challenging as it may not issue warnings to the area near the epicenter in time. The onsite earthquake early warning can compensate for its shortcomings. The peak initial-displacement amplitude (Pd) within seconds after the P-wave arrival time has been regarded as an indicator for onsite earthquake early warning. This method analyzes whether Pd exceeds a certain threshold to determine if an alarm should be issued. However, relying on a single threshold in the Pd method to predict future ground motions increases uncertainty due to the high complexity of seismic signals. It makes the Pd method a nonlinear problem. Machine learning is well-suited for solving nonlinear problems. Previous studies have proposed using Long Short-Term Memory (LSTM) neural networks to analyze signals within a short time window after the P-wave arrival time. They utilize a PGA of 80 gal as the criterion for issuing earthquake alerts. By employing neural network models to output alert probabilities, they determine whether an alert should be issued. Their results have effectively improved the performance of onsite earthquake early warning systems. In previous studies, it was pointed out that using the PGA as the criterion for issuing earthquake alerts could lead to higher instability due to the possibility of anomalous high values. As a result, it may not adequately reflect the actual disaster situation. Instead, the PGV is considered a more suitable indicator for disaster assessment as it correlates more closely with the severity of the disaster. In recent years, attention mechanisms have been widely used in neural networks for natural language processing tasks due to their ability to capture long-term dependencies and enable parallel computation. Attention-based models have shown superior performance compared to LSTM in many tasks. This study will construct a model based on attention mechanisms for onsite earthquake early warning. It will use short-time window acceleration and velocity signals, as well as station corrections, as input data, and the model will provide the probability of reaching PGV of 15 cm/s. Four recent significant earthquake events that occurred in Taiwan were selected to test the model performance, and a comparison was made using LSTM model and Pd method as benchmarks. The results show that the attention-based model in the region within 70 kilometers of the epicenter has a false alarm rate of 30.3%, a miss alarm rate of 7.4%, an F1 score of 69.4%, and provides an average of 6.35 seconds of lead time.