The multi-kernel support vector machine optimized by a genetic simulated annealing algorithm is proposed to effectively identify complex fault characters for the gearbox fault diagnosis. Fault vibration signals are processed by empirical mode decomposition to obtain several stationary intrinsic mode functions. Then, the instantaneous amplitude energy of the intrinsic mode functions are computed and regarded as the input characteristic vector of the multi-kernel support vector machine optimized by the genetic simulated annealing algorithm for fault classification. The multi-kernel support vector machine can fit an arbitrary function within a high-dimensional feature space. The genetic simulated annealing algorithm demonstrates a superior performance of global optimization and convergence speed, and can improve the diagnosis performance and robustness of the gearbox fault diagnosis model. The gearbox fault diagnosis experiment thus demonstrates the effectiveness of this novel method.