Aiming at the problem that it is difficult to extract fault features from the nonlinear and non-stationary vibration signals of steam turbine bearings, which results in low diagnosis and recognition rate, a feature extraction method based on MIGA-VMD decomposition and t-SNE was proposed. The decomposition effect of variational mode decomposition (VMD) in bearing fault diagnosis largely depends on the number of decompositions and the selection of penalty parameters. In order to achieve the optimal selection of VMD related parameters, multiple-island genetic algorithm was used to optimize the VMD parameters. The parameter-optimized VMD was used to decompose the original vibration signal of the bearing into a number of intrinsic modal functions, and the permutation entropy of the partial modal components with higher correlation with the original signal was calculated to form the fault feature.t-SNE method was used to reduce the dimension to obtain low-dimensional feature vector and it was input to support vector machine classifier to realize fault type diagnosis. This method was applied to bearing fault diagnosis and compared with four feature extraction methods: EMD+ permutation entropy+t-SNE, EEMD+permutation entropy+t-SNE, LMD+permutation entropy+t-SNE, traditional VMD+permutation entropy+t-SNE. Experimental results show that this method can be used to more accurately extract the fault characteristics of the bearing, and to effectively realize the bearing fault diagnosis.