For the nonlinear and non-stationary characteristics of rolling bearing signals, reasonable feature selection can improve the fault diagnosis rate. A fault diagnosis model based on multiscale permutation entropy (MPE) and improved whale algorithm (IWOA) was proposed to optimize support vector machine (SVM). Firstly, signal denoising was preprocessed by VMD, and multi-scale permutation entropy was calculated to reconstruct signal features. Secondly, inertial dynamic weights were introduced to improve the whale algorithm, and SVM parameters were trained to establish the IWOA-SVM fault diagnosis model. Finally, the bearing data set of Case Western Reserve University was used for simulation experiments. The results show that, compared with multi-scale entropy, MPE can represent more abundant feature information, and the fault recognition rate is improved by 2.1%. Compared with other optimization algorithms, the fault diagnosis model optimized by IWOA based on SVM has fast convergence speed, short training time and high fault recognition accuracy, which can effectively diagnose rolling bearings.