To monitor the health of rolling bearing, the vibration signals are always used for fault diagnosis. However, the non-linear and non-stationary characteristics of vibration signals have not been solved in current methods. In this work, an intelligent fault diagnosis method is proposed, which is a sequential combinations of variational mode decomposition（VMD）, Kurtogram, and artificial fish algorithm（AFSA）. To begin, original vibration signals are decomposed into intrinsic mode functions（IMFs） using VMD, among which the most effective fault information is selected based on the Kurtogram algorithm and the rules of maximum correlation coefficients. Then the feature vectors are identified using the morphological entropy and energy entropy of the above IMFs. Next, two crucial tunable parameters, penalty coefficient C and Gaussian kernel width coefficient σ are optimized through AFSA algorithm. At last, the fault diagnosis model is developed based on AFSA-SVM algorithm, in which the extracted fault features are employed as inputs. The experimental results show that the proposed method accurately identifies fault features of the original signal. It has also improved model learning efficiency and classification accuracy.