A fault diagnosis method based on the least square support vector mach ines (LSSVM) and the simulated annealing algorithm was proposed. Better paramete rs of the regularizing variable λ and the kernel width σ were obtained by usin g the simulated annealing algorithm, and the sensitive subset of features was de termined simultaneously. To verify the effectiveness of the method, roller beari ngs were tested under four operating conditions, five different shaft speeds and two load levels, and 52 features were extracted from the bearing vibration sign als. The results show that the method has a higher accuracy of classification fo r bearings fault than other methods, and it is a promising approach to condition monitoring of rotating machinery.