人工智慧技術建立微分區地震預警系統相關研究(II);Artificial intelligence technology to establish next early warning system

NCU Institutional Repository > 地球科學學院 > 地球科學學系 > 研究計畫 > Item 987654321/88707

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/88707

題名:	人工智慧技術建立微分區地震預警系統相關研究(II);Artificial intelligence technology to establish next early warning system
作者:	馬國鳳;黃信樺;吳逸民
貢獻者:	國立中央大學地球科學學系
關鍵詞:	人工智慧;地震預警系統;;Artificial intelligence
日期:	2022-07-26
上傳時間:	2022-07-27 10:46:31 (UTC+8)
出版者:	交通部中央氣象局
摘要:	本計畫之研究目的即運用大數據深度學習法於地震預警系統中，期望能快速地解算各地區最大震度並提供較長的預警時間。上一年度(2021)以“轉換器地震警示模式”(Transformer Earthquake Alerting Model, 簡稱TEAM)模型為架構，使用2012-2020年中央氣象局TSMIP的地震資料進行深度學習與測試。第一期研究成果顯示台灣地區的TEAM模型僅需4秒的地震資料即可收斂得到一相對準確的PGA預估值。今年度(2022)將對目前使用的TEAM模型做更嚴謹的檢視與進一步的調整，例如：優化資料前處理、小地震事件下採樣及加入地震規模較大之地區的資料集等，以探討前處理對於地震度預測結果之影響及使得TEAM模型更加本地化。區域型地震預警系統—即利用最早收到地震P波訊號的少數測站對地震規模及後續較具災害性之S波震幅大小進行快速的預判—是現今地震防、減災最主要與唯一可行的方法。其過程包括震源參數的解算(如規模、位置)與震度(即地動程度)的預估兩部分。台灣現行區域型地震預警系統已然成效卓著。對於島內的地震平均發報時間約在16秒左右；島外的地震則約23秒。資料傳輸時間在現今資訊硬體設備的進步與更新下已壓縮到最低，主要耗費的時間仍在初期穩定的震源參數解算上。另一方面，近年數個中大型地震如2016年美濃地震與2018年花蓮地震皆展現強烈的破裂方向性，造成局部地區超出預期的震度與破壞。由於現行預警發報準則取決於震度的大小，此誤差將直接導致部分地區預警的盲區。因此，如何在最快的時間裡解算出地震的規模、位置與破裂方向性是能否讓地震預警系統在時效與準確度上進一步提升的關鍵。本研究將利用改進與加速定位方法、整合即時破裂方向估算與建立具方向性之強地動衰減式等方面著手，提升地震預警系統的效能與減低預警誤差與盲區的發生。近年來中央氣象局逐步提升即時觀測站傳輸效能與密度，進而使整體觀測系統在地震定位時能更加快速、準確。另一方面，由國立臺灣大學團隊維護的高密度地震觀測網—P-Alert觀測網，整合了區域型預警、現地型預警、地震速報以及提供高品質地震紀錄等功能。為測試高密度觀測網在地震預警的效能及優勢，目前已初步完成兩系統之合併測試，結果發現定位及規模決定(tcpd)與即時波形撿拾(pick_eew)等流程都存在潛在問題，亦影響了輸出結果的品質。目前已針對各個潛在問題提出初步解決方案，預期在修正相關潛在問題後，能產出良好且穩定的結果。預期此整合型系統最終應能對地震防、減災以及地震科學研究產生極大的貢獻。 ;The research purpose of this project is to use deep learning method in earthquake early warning system for quickly estimating peak ground acceleration (PGA) in any site and also for providing a longer early warning time. In 2021, the CWB earthquake catalogs from 2012 to 2020 were trained and tested under the deep learning structure of Transformer Earthquake Alerting Model (TEAM). The first phase of the research results showed that the TEAM model in Taiwan only needs 4 sec of waveforms to get a relatively accurate value of predicted PGA. This year in 2022, we will focus on improving the currently used TEAM model by optimizing data pre-processing, undersampling of small earthquake events, and adding dataset from regions with larger earthquakes. This rigorous review and adjustment of TEAM model can help us understand how the attributes in pre-processing affect the predicted PGA intensity results and can make TEAM more suitable for Taiwan. Regional earthquake early warning system (EEWS), which utilizes P-wave signals recorded by first few stations to rapidly estimate the earthquake magnitude and upcoming large S-wave shaking, is the primary means to mitigate earthquake hazards nowadays. Two main components of the system are estimation of source parameters (e.g. magnitude and location) and prediction of intensity (i.e. shaking levels). Currently, Taiwan EEWS has made great achievements in the world and can issue reports within 16 s and 23 s for inland and offshore events, respectively. The report time takes mostly in estimating reliable source parameters rather than data transmission as the advance of hardware and telemetry techniques has reduced the transmission time at most. Moreover, a number of recent earthquakes with ML>6 produced directional strong shaking and unexpected damage. Since the current criteria of issuing EEWS report are based on intensity, this underestimation of the strong shaking has a direct effect on blind zone formation. To further improve the efficiency of reporting time and accuracy of intensity prediction, how fast to obtain the source parameters including rupture directivity is the key. We plan to improve earthquake location procedure, incorporate rupture directivity at near-real time basis, and establish directivity-guided ground motion prediction equations to help increase the EEWS efficiency and reduce ground motion prediction uncertainty and blind zones. Recently, the Central Weather Bureau (CWB) has gradually improved the transmission efficiency and density of real-time seismic stations, thereby making the overall seismic monitoring system faster and more accurate in earthquake locating. On the other hand, the P-Alert network, a high-density seismic monitoring network maintained by the National Taiwan University research team, integrates functions such as regional and on-site earthquake early warning, earthquake rapid report and providing high-quality earthquake records. In order to test the effectiveness and advantages of the high-density seismic network in earthquake early warning, a preliminary combined testing of the two systems has been completed. The results found that the location and magnitude determination (tcpd) and real-time waveform picking (pick_eew) processes have some potential problems, which have also lowered down the quality of the result. At present, preliminary solutions have been proposed for each potential problem, and it is expected that after correcting the related problems, fine and robust results can be produced. It is expected that this integrated system could make a great contribution to earthquake disaster prevention, mitigation and earthquake scientific research.
關聯:	財團法人國家實驗研究院科技政策研究與資訊中心
顯示於類別:	[地球科學學系] 研究計畫

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	92	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....