建物震害毀損度預測模式之研究－ 倒傳遞類神經網路法之應用A Forecast of Building Destruction in Earthquakes: Applications of Artificial Neural Network

本研究係以文獻回顧之相關研究、921地震時七級震度之竹山地區建物毀損資料及現有建物調查之資料，尋求影響因子與變數，並將上述資料以倒傳遞類神經網路MATLAB6.5軟體，利用其具有學習及記憶能力加以訓練、測試及驗證，建立中低層建物震害毀損度預測模式，應用於嘉義市部分舊市區，並以地理資訊系統予以空間化，將建物毀損度分為安全、危險與倒塌三級，結果證明類神經網路具有預測建物震害毀損度之能力，且其誤判率較低，是為都市防災規劃值得推廣及應用的方法。
關鍵詞：倒傳遞類神經網路、建物震害毀損度、誤判率

This research investigates potential variables of buding destruction in earthquakes and their influences by conducting an empirical study of Jwu-Shan area in Taiwan. Jwu-Shan was one of the most serious damaged areas during the 921 earthquake which was 7.0 magnitude. In this study , MATLAB6.5 software of back-propagation neural network was used with its superior attributes, i.e., learning and memory, to establish a forecast model of hazards in middle and lower buildings in an earthquake by means of training, testing and validation. The model was tested with the data of partial old communities in Chia-Yi. Damaged buildings were classified into 3 categories: safe, unsafe, and collapse by geography information system (GIS) with data spatialization and transformation. The results suggest that the artificial neural network is capable to forecast building destruction in earthquakes with a low error rate. The paper concludes with applications of a back-propagation neural network in planning urban disaster prevention.
Key words: back-propagation neural network, building destruction in earthquakes, error rate