In order to extract the gear fault features, firstly, the gearbox vibr ation signal was decomposed as intrinsic model functions (IMF) by using the empi rical mode decomposition (EMD) method. The energy of every IMF and the singular value of the IMF matrix were chosen as features. The Shannon and Renyi entropy o f the energy and singular value distribution were also extracted. Secondly, a wr apper feature selection method employing the genetic algorithm and the least squ are support vector machine (LSSVM) was used to search the optimal feature subs e ts for the gear fault diagnosis. The results demonstrate that the proposed appro ach can detect the gear faults by only using a small feature set with high accur acy and efficiency.