應用多層函數連結網路預測深開挖壁體變位之研究

Prediction of Diaphragm Wall Deflection in Deep Excavations by Multilayer Functional-Link Network

吳建逸(Chien-Yi Wu) ； 陳俶季(Shuh-Gi Chern)

《中國土木水利工程學刊》 24卷1期 (2012/03) Pp. 1-13

https://doi.org/10.6652/JoCICHE.201203_24(1).0001

摘要

基礎深開挖在設計與施工前，為求施工安全、周邊建築物保全及其重要公共設施保護等原因，一般會依據地層條件及外在環境因素等分析數據，透過電腦程式模擬演算開挖後擋土壁體變位及其最大變位量和所發生位置等資訊，再由此數據分別訂定出安全管理值。如在開挖階段發生非預期變位較大時，可適時提醒現場施工管理者應變，並能有足夠時間採必要之補強措施。故基礎開挖階段壁體變位量之預測就極為重要。本研究搜集整理17個案例，得訓練範例數2,475筆，測試範例141筆，運用多層函數連結網路（MFLN）之訓練與測試學習功能，建立「類神經網路深開挖壁體變位預測系統」。根據測試與驗證結果顯示，逐項所建構之多層函數連結網路預測模式，不管在每一觀測點變位值、最大壁體變位值及其發生位置預測上，皆有不錯的預測結果。與BPN網路及RIDO程式預測準確度相比，有所提升約5％以上。另應用在同地區其他案例中，經預測驗證結果顯示，亦同有良好預測準確度，代表本研究預測系統運用在該地區之適宜性。

關鍵字

深開挖 ； 壁體變位 ； 多層函數連結網路 ； 預測系統

並列摘要

Before design and construction of deep excavation, for the sake of construction safety and protection of adjacent buildings, prediction and measurement of diaphragm wall deflection are very important so as to avoid failures of the supporting system. This paper attempts to predict the diaphragm wall deflection in deep excavation by using multilayer function-link neural network (MFLN) learning model. 17 case histories of deep excavation with a total 2,616 sets of wall deflection from construction projects are collected. These 2,616 sets of wall deflection are randomly divided into two groups, training group has 2,475 sets and testing group has141 sets of data. For comparison, lack-propagation network (BPN) and numerical analysis model RIDO are also applied to evaluate diaphragm deflections. From the results of this research, it is shown that the MFLN model can reasonably predict both magnitude and location of the maximum deflection of the braced wall in deep excavation.

並列關鍵字

Deep excavation ； Wall deflection ； Multilayer function-link neural network ； Projects the system