A rolling bearing fault feature extraction method based on improved VMD and entropy feature fusion was proposed to address the problem of low fault diagnosis recognition rate caused by the difficulty in effectively decomposing vibration signals due to the manually selected penalty factor and the number of modes in variational mode decomposition (VMD). Firstly, the improved walrus optimization algorithm (IWaOA) was obtained by using cubic chaotic mapping and the golden sine strategy. Then, IWaOA was used to optimize VMD and find the optimal penalty factor and the number of modes. Then, VMD was used to decompose the vibration signal and calculate the seven entropy value features for each mode component, and establish an IWaOA-VMD feature extraction model. Secondly, the linear discriminant analysis (LDA) method was used to reduce the dimensional fusion of the seven entropy value features, and the fused feature vectors were input into the back propagation (BP) neural network for identification, so as to established the LDA-BP fault identification model. Finally, the proposed method was verified by experiments. The results show that the fault recognition accuracy of the proposed method on the Case Western Reserve University bearing data set is 99.58%, and the accuracy under strong noise interference is more than 92%. In order to verify its applicability, the accuracy of fault identification on the XJTU-SY data set of Xi'an Jiaotong University reaches 100%, which verifies the noise robustness and multi-source data applicability of the proposed method.