Kernel method is an effective approach to solve the nonlinear pattern recognition problems in the field of machine learning. At present, multiple kernel method has become a new research focus. Compared with the traditional single kernel method, multiple kernel method is more flexible, more interpretable and has better generalization performance when dealing with heterogeneous, irregular and non-flat distribution samples. A multi-kernel S³VM optimization model based on L_p norm constraint is presented in this paper in accordance with kernel method of supervised learning. Such model has two sets of parameters including decision functions f_m in reproducing kernel Hilbert space and weighted kernel combination coefficients, and inherits the non-smooth and non-convex properties from single-kernel based S³VM. A two-layer optimization procedure is adopted to optimize these two groups of parameters, and an improved Quasi-Newton method named subBFGS as well as a local search algorithm based on label switching in pair are used to solve non-smooth and non-convex problems respectively with respect to f_m. Base kernels and manifold kernels are added into the multi-kernel framework to exploit the geometric properties of the data. Experimental results show that the proposed algorithm is effective and has excellent generation performance.