In this paper, a novel bearing fault diagnosis method based on multi-layer extreme learning machine (MELM) optimized by the novel ant lion algorithm (NALO) is proposed. First, using permutation entropy of different scales (MPE) to extract fault features of bearings, a group of fault feature vectors composed of permutation entropy is obtained. Then, the fault feature vectors are classified by the MELM. However, with the increase of the number of hidden layers, the random input weight and bias will also increase, which will lead to the increase of the randomness of the MELM and affect the accuracy of fault diagnosis. Therefore, this paper uses the NALO to optimize the MELM. For the NALO, opposite populations are added to the initial population to improve its global search ability. When the ant lion updates its location, the influence of pheromones left by other ants with a certain sensing distance is taken into account to prevent the ant lion from falling into the local optimal and increased the robustness. Finally, the NALO-MELM and other bearing fault diagnosis methods are applied to the bearing fault experiment of Western Reserve University to test the performance and generalization of the proposed method.