A transformer winding core mechanical fault diagnosis method based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) multi-scale fuzzy entropy (MFE) and multi-verse optimizer algorithm optimized kernel based extreme learning machine (MVO-KELM) is proposed for the problem of insufficient accuracy of transformer winding core mechanical fault diagnosis. Firstly, to avoid the generation of spurious modal components, the original transformer vibration signal is decomposed using a modified ICEEMDAN. Secondly, the modal component with the highest correlation is selected using the Pearson correlation coefficient method and its MFE value is calculated. Then, the MFE values are used as feature quantities to construct feature datasets, and the kernel parameters and regularization coefficients of KELM are optimized using MVO. Finally, the feature dataset is input into the constructed MVO-KELM model for classification and identification to achieve the goal of high accuracy diagnosis. The experimental results prove that the proposed approach possesses excellent diagnostic accuracy and stability, and can accurately diagnose transformer winding core loosening faults of different degrees with a diagnostic accuracy of more than 99%, which may supply the necessary guidance for the development of transformer field maintenance strategy.