A hybrid approach combining uniform design and support vector machine to probabilistic tunnel stability assessment

Hybridization of UD and SVM proposed in RSM context for tunnel-reliability analysis.Use of merits of both SVM and UD made to handle implicit LSF for tunnel-reliability.Computational accuracy and efficiency demonstrated by illustrative examples.Application potential verified for tunnel-reliability problem posed by implicit LSF.

Applications of the reliability-based method to stability evaluation of tunnel structures have become an ever-increasing concern over recent years. One critical challenge in conducting such a task is the implicit nature of the limit state function (LSF). To address this issue, the focus of this study is on, among others, the use of response surface method (RSM) by considering both the selection of the sampling method and the choice of the response surface form (as two major factors affecting the RSM’s performance). In this context, the current paper develops for tunnel-reliability analysis a hybrid approach combing an experimental design called uniform design (UD) and a regression device known as support vector machine (SVM). For the proposed hybrid approach, the UD is used to generate sampling points and then the SVM is employed to construct the response surface approximating the original inexplicit LSF. Such an approach integrates the merits of both UD and SVM used for complex nonlinear modelling. Three carefully selected tunnel examples are illustrated: one for a typical tunnel under relatively simplified tunnelling conditions and the other two for real-life tunnels. Comparisons are made to validate the computational accuracy and efficiency of the present approach. In particular, for the tunnel example where the LSF is known only implicitly through the numerical analyses (which is the scenario of many real-world applications in tunnel community), the obtained results further demonstrate the efficiency of this approach: it can be much more economical to achieve reasonable accuracy than the conventional RSMs when a small number of sampling data is used. Such comparisons made in this work verify the application potential of the developed hybrid approach for probabilistic tunnel stability assessment involving the implicit LSF.