Computational probabilistic methods enable us to incorporate and propagate uncertainties in mechanical models. However, in some cases, classical methods, such as FORM/SORM methods and Monte-Carlo methods, can be computationally expensive or inaccurate. An efficient importance sampling method is then suggested to yield sufficiently accurate results with acceptable computational cost in an industrial context. The method is an importance sampling method based on a second order asymptotic approximation combined with the HyperCube Latin method. A clustering method is used to solve the global optimization problem which arises to find the points of maximum likelihood. The efficiency of the method compared to classical methods is illustrated with several examples. Considerable reduction of the statistical error of the estimated failure probability can be achieved. The interest of the method is assured provided the points of local maximum likelihood are not too numerous and uniformly distributed. The paper presents two vibratory test cases, the second one is an industrial piping system.