| 摘要: | 台灣位處環太平洋地震帶，地震活動頻繁，城市地區因老舊建築比例增加與高人口密度，使地震風險進一步加劇。然而，現行的耐震評估方法依賴大量人力與時間，難以快速應用於城市規模建築檢測，不利於即時防災與決策支援。因此，本研究提出一種結合深度學習與高效能運算（High Performance Computing, HPC）的機率式地震風險評估架構，專注於高地震風險區域的大量中低矮鋼筋混凝土建築，目標旨在提升城市規模評估的效率與準確性，並為政府與相關機構提供決策依據，以降低災害可能帶來的影響。
本研究首先運用YOLOv5物件偵測模型自動提取建築樓層數，透過二階段遷移學習策略，將交集比為0.5時的平均精度 (Average Precision, AP），亦即 AP_0.5 提升至91%，有效解決城市內大量建築基本資訊蒐集問題。接著，開發機率式地震風險評估框架 (FEMA P-58)，結合增量動力歷時分析考慮地震動態反應與不確定性，並利用OpenSeesMP平行運算技術與國家高速網路與計算中心 (NCHC) 計算資源，有效減少大量房屋進行地震風險評估所需之時間，最終透過易損曲線量化建築在地震下的損傷程度。
本研究最後以花蓮地區 361 棟 921 大地震前建造之 1 至 8 層樓建築，及 17 棟於 0403 花蓮地震後實際張貼紅黃單之建築為例，應用所提出之地震風險評估方法進行模擬分析，並分別與台灣地震損失評估系統TELES所推估之損傷分布，以及實際災損情形進行比對。結果顯示，本研究方法所推估之建築受損狀況與地震強度分布高度一致，亦能合理反映實際貼單情形，驗證本方法具備良好之準確性與適用性。
關鍵字：機率式地震風險評估、高效能運算、深度學習、0403花蓮地震
;Resilient cities aim to enhance post-disaster recovery. Taiwan′s urban areas face severe seismic threats, worsened by aging buildings and high population density. However, conventional seismic assessments are time-consuming and labor-intensive, limiting large-scale applications. This study proposes a probabilistic seismic risk assessment method integrating deep learning and- high-performance computation, focusing on low-to mid-rise reinforced concrete buildings in high-risk zones.
The YOLOv5 object detection model was first used to extract building floor numbers, Achieved an Average Precision (AP) of 91% at an Intersection over Union (IoU) threshold of 0.5 (i.e., AP_0.5=91%), through a two-stage transfer learning strategy. A probabilistic assessment framework (FEMAP-58) incorporates incremental dynamic analysis (IDA), which was then developed to capture the uncertainties in seismic response. Using OpenSeesMP and parallel computing developed at the National Center for High-Performance Computing (NCHC), computation time is effectively reduced for region-scale assessment. Finally, Fragility curves quantify seismic damage levels.
Last but not least, this research used 361 pre-1999 buildings in Hualien to demonstrate the application of the proposed framework at the regional scale. In addition, 17 buildings that were officially red- or yellow-tagged after the April 3, 2024, Hualien earthquake were included for further simulation. A comparison was made with the Taiwan Earthquake Loss Estimation System (TELES) results for the April 3, 2024, Hualien earthquake and actual post-earthquake damage records.. The simulated damage distribution showed a strong correlation with seismic intensity and actual building damage, validating the accuracy and applicability of the proposed method.
Keywords: Probabilistic Seismic Risk Assessment, High Performance Computing, Deep Learning, April 3, 2024, Hualien Earthquake |
