以人工智慧理論探討臺中市中小學校舍耐震因子及耐震能力

To Interpret the Seismic Factors and Seismic Abilities of School Buildings in Taichung City using Artificial Intelligent Theories

臺灣位於環太平洋地震火環帶，地震發生機率極高。臺中市位於臺灣強震區，人口密集，歷年來所發生的大地震，造成多棟中小學校舍破壞及生命財產損失。由於中小學校舍常做為地震侵襲後之避難所，地位極為重要，故本論文以校舍之耐震因子及耐震能力為研究主題。所使用的研究方法，包括:主成份分析法、資料探勘、灰色理論以及倒傳遞類神經網路。主成份分析法乃利用特徵值的觀念來歸納出耐震因子; 資料探勘則應用於校舍的分群; 灰色理論可分析校舍耐震因子與崩塌地表加速度之灰關聯係數，了解耐震因子與崩塌地表加速度之關聯程度; 再透過倒傳遞類神經網路之學習、回想、歸納推演三項主要功能，訓練及測試中小學校舍之耐震能力評估模式。實證案例則以國家地震工程研究中心所提供的中小學耐震資料庫為研究樣本，並且實際調查測繪臺中市市中心區326棟校舍，透過上述的研究方法，探討臺中市中小學校舍重要耐震課題。從結果中可知，各分群之耐震因子與崩塌地表加速度之灰關聯度排序並不相同; 各分群經類神經推論後的誤差均方根值介於0.0086至0.0829之間; 判定係數R2則介於0.6847至0.9966之間，上述推論並與支持向量機推論做比較，結果頗佳，可提供給建築師規劃設計校舍時使用，亦可供後續學術研究之參考。

Taiwan, located along the seismic belt that fringes the western Pacific at the junction between the Eurasian and Philippines Sea Plates, sees frequent seismic activity. Taichung City, a densely populated area of central Taiwan, had many school buildings damaged after earthquakes. Owing to school buildings serve important dual roles as places of education and as public shelters following earthquakes. This paper probed into the seismic factors and seismic abilities of school buildings in Taichung. The main research methods of this paper include the principal component analysis, data mining, grey theory and back propagation neural network. The concept of principal component analysis was used to generalize the seismic factors by Eigen-values. Data mining was adopted for school buildings’ clustering. Grey theory was utilized for the grey relationship between seismic factor and collapse ground acceleration. Back propagation neural network was used to deduce the seismic assessment models of the school buildings. National Earthquake Engineering Research Centre provided some seismic data of school building for this paper. This paper used these data for research specimen and investigated the seismic factors of school buildings in Taichung City. From the results know that the sequence of the grey relationship grades differs in every clustering; the RMSE values range from 0.0086 to 0.0829; the R2 values are between 0.6847 and 0.9966. The results also compared with the deduction of support vector machine. The fruitful results can provide for architects to use, also can provide academics for references.