| 摘要: | 本研究利用XGBoost模型，結合台灣地震危害高階模型建置（TWSSHAC）計畫蒐集台灣強震觀測網（TSMIP）地殼地震紀錄所得之Flatfile資料庫，分別針對PGA、PGV、以及反應頻譜建立地殼地震模型（GMM）。衡量模型性能時，本研究特別關注殘差標準差和R²分數兩項指標，及透過距離響應、反應頻譜和殘差分佈等多重分析手段，深入評估模型的有效性與合理性。此外，研究方法著重於應用可解釋AI工具SHAP（SHapley Additive exPlanations）指數來揭示不同預測因子在本研究模型中的重要性。此方法不僅闡明了各種因素對強地動的影響，更強調模型的可解釋性，滿足傳統工程地震學對於模型的合理性與可解釋性的需求。
考慮到地震數據集的不均衡問題，特別是在較大規模事件和較短距離的情況下，採用SMOGN方法進行數據增強。這種方法有效平衡數據集，從而使模型在預測大規模地震事件方面具有更好的學習能力。本研究亦評估模型在地震危害分析中的可用性，採用地震危害分析軟體Openquake，納入本研究強地動衰檢模型做地震動模擬。將機器學習技術納入GMM的發展，標誌著台灣地震危害分析的重要進展。這項研究可能對地震準備策略、基礎設施韌性規劃和公共安全協議產生重大影響，展示了機器學習在地震科學中的應用潛力。
;This study utilizes the XGBoost model integrated with the Flatfile database, derived from seismic records of the Taiwan Strong Motion Instrumentation Program (TSMIP) and archived by the Taiwan Seismic Senior Hazard Analysis Committee (TWSSHAC) project, to establish crustal seismic models (GMM) for PGA, PGV, and spectral acceleration. When evaluating the performance of these models, particular attention is paid to the standard deviation of residuals and the R² score, as well as multiple analysis methods such as distance response, response spectra, and residual distribution to thoroughly assess the effectiveness and rationality of the models. Additionally, a key aspect of the methodology is the application of SHAP (SHapley Additive exPlanations) values to reveal the importance of different predictors within our model. This approach not only clarifies the impact of various factors on strong ground motion but also highlights the interpretability of the model, addressing the traditional demands of engineering seismology for model rationality and explainability.
Considering the imbalance in the earthquake dataset, especially in cases of larger-scale events and shorter distances, the SMOGN method is used for data augmentation. This method effectively balances the dataset, thereby enhancing the model′s ability to learn from rare but significant large-scale seismic events. The study also assesses the applicability of the model in seismic hazard analysis using the seismic hazard analysis software, Openquake, incorporating the strong motion attenuation model proposed in this study for seismic motion simulation. The incorporation of machine learning techniques into the development of GMM marks a significant step forward in the advancement of seismic hazard analysis in Taiwan. This research could have a substantial impact on earthquake preparedness strategies, infrastructure resilience planning, and public safety protocols, demonstrating the potential of machine learning in earthquake science. |
