- 學位論文
以類神經網路觀察近斷層之地震波方向性
The Observation of Seismic Wave Directivity in Near-Fault Zone with Deep Neural Network
摘要
在過去的研究中發現，地震波會在某些特定的條件或因素下，會呈現明顯的方向性，如:當地剪力波速的異質性(Bonamassa & Vidale, 1991)、山脊方向的放大現象(Spudich et al., 1996)、地層破裂帶的折射反應(Xu et al., 1996)和斷層附近震波的偏震現象(Lombardo, 2008)，這些造成地震波有方向性的因素對設定地震參數(GMPE, Ground Motion Prediction Equation)有深遠的影響，並間接影響到一般建築物耐震設計規範，現行的建築物耐震設計規範對近斷層效應只考慮垂直於斷層方向的放大現象，，而過去的研究發現地震波的方向並不會完全平行於斷層的垂直方向，因此規範對此現象還沒有完整的考量，本研究將針對近斷層場址內的測站進行地震波方向性的分析。 本研究目前蒐集了全臺灣104場地震事件以及基本之地質資料。造成近斷層地震波產生方向性的原因主要可以分成三類:震波偏振、剪力波速下降以及向破裂前進方向效應(forward directivity)，另外地震的來源也有可能會影響地震波的方向性，根據這些因素選出地震、斷層和測站相關之因子，並利用統計以及類神經網路的方法探討各因子對地震波方向性之影響。本研究採用愛氏強度(Arias Intensity)來評估地震波之方向性，並嘗試利用類神經網路來預測在各種近斷層條件下的愛氏強度分佈情形。 總結統計方法的結果，主要的影響因子包括場址和震央之距離、地震震源深度、隱沒帶地震類別、場址地層主頻。在類神經網路訓練階段將重要因子一一抽離，確實會造成類神經網路的表現受到影響。而由類神經網路之權重的結果發現，現地應力狀態、斷層走向以及斷層滑動機制皆為重要之影響因子。 由本研究採用之類神經網路之訓練結果顯示，大部分近斷層場址之地震波會被歸類無方向性，且對於沒有方向性之地震波資料可以有八成以上之預測準確率，但對於有方向性之地震波資料卻只有四成的準確率。雖然在真實情況下無方向資料確實佔大多數，但有方向性資料之準確率無法再提升，可能顯示本研究所採用之參數還不足夠描述震波之方向性。
並列摘要
Based on previous research, ground motion can be amplified in certain direction and show with significant anisotropy. The causes still remain unclear, and different researchers have attributed this phenomenon to several factors, including topographic effect, local geological heterogeneities, wave polarization, wave trapped in fault zone and etc. This phenomenon might have severe impacts on buildings that cause damages, especially in the near-fault area. However, the current seismic design code focus on the perpendicular direction of fault strike only, which is not suitable enough for real situation. The objective of this study will focus on seismic wave directivity in near-fault zone. A total of 104 earthquake events with basic geological data were collected. Causative factors were selected based on previous research. There are three main causes considered of free field stations, included wave polarization, anisotropic stiffness and forward directivity. The source of earthquake can possibly affect the seismic wave directivity. Data of influence factors were collected accordingly, and Arias Intensity is used to describe the directivity of seismic wave. The deep learning technique was applied to predict Arias Intensity distribution with the given parameters. This research used TensorFlow as the main deep learning tool. The results of neural network shows that, in most cases, there were no obvious directivity in the near-fault site, and the accuracy of non-directional data was higher than the accuracy of directional data. Although, in real situation, non-directional data is the majority, the low accuracy of directional data might suggest the parameters used in this research is not enough to describe seismic wave directivity. The results suggested that the distance between site and epicenter, focal depth and subduction earthquake are all important factors. In addition, neural network takes in-situ stress, fault strike and fault slip type as main factors. This result will be discussed in this paper.
參考文獻
延伸閱讀
- 陳嘉榮（2013）。利用震波波速變化 辨識斷層於近地表位置之可行性研究〔碩士論文，國立中正大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0033-2110201613565852
- 林鈺棠（2011）。以強震數據與地層特性資料建構類神經網路模式之研究〔碩士論文，國立屏東科技大學〕。華藝線上圖書館。https://doi.org/10.6346/NPUST.2011.00259
- 施啟揚（2014）。運用類神經網路於中間層隔震建築物之系統識別〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU.2014.01410
- 游昌梓（2004）。斷層之摩擦組成律與地震能量釋放型態初步探討〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU.2004.01903
- 陳美蓮（2009）。Numerical Modeling of Wave Deformations in the Nearshore Zone〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU.2009.01909