不確定性(uncertainty)普遍存在於大地工程中,在可靠度設計中是一個重要的指標。目前業界使用的安全係數設計法雖然具有方便且迅速的優點,但無法準確量化不確定性因子,可能導致過於保守的設計。因此本研究的目的即是有效利用現場或室內試驗的資訊來預測岩體之變形模數的機率分佈情形,並結合不僅僅單一參數的資訊來降低其不確定性。首先,藉由文獻回顧蒐集前人對岩石經由現地或室內試驗所得之岩石參數資料,建立一龐大的資料庫,篩選出我們認為與岩石變形性有密切關聯的參數,包含:(1)RQD;(2)RMR;(3)Q-System;(4)GSI;(5)完整岩石之楊氏模數(Young's modulus of intact rock, Er);(6)完整岩石之單軸抗壓強度(uniaxial compressive strength of intact rock, σc),並利用這六種參數來預測岩體的變形性參數,包含:(1)岩體之變形模數(deformation modulus of rock mass, Em);(2)岩體之彈性模數(elasticity modulus of rock mass, Ee);(3)岩體之動態模數(dynamic modulus of rock mass, Edyn)。接著,應用貝氏機器學習(Bayesian machine learning)於此資料庫,建立通用的多變數機率分佈模型,並且量化資料庫中的空洞所造成的統計不確定性。根據此多變數機率分佈模型,可以得出參數間的相關性,並且能在不同參數組合的條件下,對岩體變形模數進行預測。當輸入的已知資訊越多,預測參數的不確定性也會隨之縮小,於可靠度設計的架構下,能更準確地設計結構物,適度地節省工程材料成本。
Comparing with safety factor method, reliability-based design method can quantify the uncertainty to design geotechnical structure in a more systematical and economical design. In this study, a multivariate probability distribution model for nine parameters of rock is constructed based on the RM/9/5890 database by a Bayesian machine learning method. These nine parameters are: (1) RQD; (2) RMR; (3) Q-System; (4) GSI; (5) deformation modulus of rock mass (Em); (6) elasticity modulus of rock mass (Ee); (7) dynamic modulus of rock mass (Edyn); (8) Young's modulus of intact rock (Er); (9) uniaxial compressive strength of intact rock (σc). This method admits incomplete multivariate data, so it can handle missing data in the database. It can rigorously quantify transformation and statistical uncertainties. From the results, the transformation uncertainty can be effectively reduced as the multivariate site-specific information increases. With smaller uncertainty, reliability-based design can be more economical.