In order to accurately and effectively diagnose bearing fault and its fault degree under the variable load conditions, variable mode decomposition (VMD) based on improved chaotic fruit fly optimization algorithm and multi-class relevance vector machine (MRVM) based on nested one to one algorithm is proposed. Firstly, the improved chaos fruit fly optimization algorithm(ICFOA) is used to optimize the penalty parameter and component numbers of intrinsic mode function (IMF) of the VMD, and search for the optimal combination of two parameters. Next, the key parameters of VMD are set by using the optimal combination parameter value, and the corresponding IMF components are obtained by decomposing the known fault signals. Then, the nested one-to-one algorithm is used to construct a high-precision multi-classification RVM learning model, and the 2-D marginal spectral entropy of IMF components are used as the input feature vector of MRVM. Finally, the experimental data under different loads are used to verify the results. The experimental results show that the proposed method can accurately diagnose bearing faults under variable load conditions, in which the diagnostic accuracy of bearing fault type is 100%, the diagnostic accuracy of bearing fault degree is 91.87%, and with high accuracy and robustness.