Abstract: Reliability analysis of the anti-sliding stability of complex foundations of gravity dams is faced with a large number of random variables and a strong nonlinearity in implicit performance functions, causing a convergence difficulty with the traditional response surface method and its low accuracy and high computational cost. To solve the problems, we develop a dynamic response surface method of partial least squares coupled with extreme learning machine (PLS-ELM) to estimate the reliability index of foundation anti-sliding stability. A response surface of implicit performance functions is constructed using ELM that is suitable for solving nonlinear regression problems of high-dimensional small samples. This surface is combined with the Monte Carlo simulation method through optimizing the number and input weights of hidden layer neurons with the PLS technique, and its dynamic update and the solution of reliability index are realized by constructing a reasonable iterative algorithm. Using this new dynamic method, we calculate the reliability index of foundation anti-sliding stability for a real dam. The results show its advantages of low computational cost, high accuracy and efficiency in response surface fitting, easy convergence, wide applicability, and easy integration with various methods or software for structural calculation. Thus, the method is suitable for the reliability analysis in the design of real projects. A case study demonstrates that a reliability index of 5.10 meets the standard requirements and the material parameters of soft interlayers have the greatest influence on the reliability of dam foundation anti-slide stability.

Key words: hydraulic structure, foundation anti-sliding stability, reliability, extreme learning machine, partial least squares, dynamic response surface