| 摘要: | 本研究提出一基於深度學習的端對端（end to end）區域型地震預警模型
「Taiwan Transformer Shaking Alert Model（TT-SAM）」。此模型考慮數個測站之地震波形，透過卷積神經網路（CNN）進行特徵擷取，結合受P 波觸發和目標測站位置和場址效應，在P 波抵達第一個測站後3 秒開始提供以PGA 為標準之預估震度圖。藉由調整靠近震央的地震紀錄在訓練流程中的權重和考慮場址效應，成功減緩高震度的低估問題。為探討此模型應用於台灣之可行性，以中央氣象署震度4 級（PGA>25cm/s2）作為預警標準，在測試集2016 年地震事件的預估表現中，精確率和召回率分別為75%和81%。在不同的P 波抵達第一個測站後之秒數，預估的精確率相近，而召回率隨著秒數增加而提升，在透過滾動式預警計算預警時間中，2016 年兩個規模 6.0 以上的美濃與台東外海地震，平均預警時間分別為16 秒和7 秒。為確認模型的可信任性，除了分析預估誤差的空間分佈，也分析CNN 特徵圖與數個震波物理參數之相關性，其中以垂直分量波形包絡線的相關性最高。除此之外，透過評估本文模型與氣象署預警系統之效能，提供預估震度與預警時間之比較結果。綜合以上，透過優化TT-SAM 架構，運用台灣地震資料重新訓練，其優勢為能夠快速且有效預估震度4級以上區域，並在後續提供模型解釋，以協助評估模型的可信任度。期望未來能透過持續引入更新的高震度資料微調模型以改善高震度低估問題，與將其架構延伸至預估PGV，以提供與實際災損分佈更相符之地震預警。;This study has referenced the Transformer Earthquake Alerting Model (TEAM), a deep learning earthquake early warning (EEW) framework. I optimized the model using seismic data from Taiwan to develop the Taiwan Transformer Shaking Alert Model (TT-SAM), and it could rapidly calculate the seismic intensity to provide longer warning time. The model utilized the Taiwan Strong Motion Instrumentation Program (TSMIP) database to obtain waveforms for events with a magnitude greater than 3.5 that occurred between 1999 and 2019. I split the dataset for model training and testing, the observations in 2016 were separated individually for the final evaluation. I cut the waveform initially triggered by the P-wave into a time window of 15 seconds, and other triggered stations′ waveforms in these 15 seconds will also be included. The model extracts waveform features through a convolution neural network (CNN), while the transformer encoder builds the relationship between features and station location. At the end of the model, a mixture density network was implemented to predict ground shaking by probability density functions. A warning threshold at 25 cm/s2 in PGA was set, corresponding to Central Weather Administration (CWA) intensity IV, to validate the model′s performance with 2016 data. The result shows that precision and recall are 75% and 81%, respectively. While utilizing the rolling warning method, it′s noteworthy that the average lead time for the 2016 Mw6.4 Meinong event and Mw6.1 Taitung offshore event stand at 16 and 7 seconds, respectively. In order to validate model’s reliability, I analyzed not only the residual of predicted PGA at different station also the correlation between the feature map from CNN and other waveform physical attributes, e.g., waveform envelope. The objective is to improve the efficiency and dependability of AI-based EEW in Taiwan. In addition, evaluating the performance of the TT-SAM model and the CWA EEW system provides comparative results of estimated seismic intensity and warning time. In summary, by optimizing the TT-SAM framework and retraining it with Taiwan′s earthquake data, the model can quickly and effectively predict areas with a seismic intensity of 4 and above. Additionally, it provides model explanations to help assess the model′s reliability. In the future, the goal is to continually fine-tune the model with updated high-intensity data to address the issue of high intensity underestimation. The TT-SAM framework could also extend to predict PGV, thereby offering earthquake warnings that more closely align with actual damage distribution. |
