Abstract: To treat the vibration faults with non-stationary and multi-component features appearing in the running process of steam turbines, a new method for vibration fault diagnosis of steam turbines was proposed based on relative entropy (RE) cloud model of variational mode decomposition (VMD) and optimized least squares support vector machine (LSSVM). Firstly, the fault signal was decomposed into K modal components according to the preset scale by using the variational mode decomposition. The pseudo-components were removed according to the relative entropy of each modal component and the original signal, and subsequently optimal signal components were extracted and put into the cloud model, while the feature vectors were extracted with inverse cloud generator. Then, the improved fruit fly optimization algorithm was used to dynamically adjust the search step to find the best combination of the super-parameters that would affect the identification accuracy of LSSVM. Finally, the LSSVM with optimized input parameters of eigenvectors was used to identify the faults, and the identification results were compared with that of the LSSVM algorithm respectively based on empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) relative entropy cloud model. Results show that the proposed method is superior to traditional signal decomposition methods, which has a high recognition rate for turbine vibration faults.