To solve the problems of low computational efficiency and missing amplitude information existing in the current multiscale sample entropy（MSE） method when extracting features of complex series, refined improved multiscale fast sample entropy（RIMFSE） is presented. Firstly, fast sample entropy is employed to substitute traditional sample entropy, and the calculation cost is greatly reduced by improving the matching mechanism of reconstructed vectors. After that, the improved multiscale expansion method is applied to replace the traditional coarse-grained method, thereby avoiding the loss of amplitude information. Based on this, a new rotating machinery fault diagnosis method is proposed in combination with the max-relevance and min-redundancy（mRMR） method and the support vector machine（SVM） classifier. Two fault data sets of gearbox and bearing are used to verify the performance of the presented method; meanwhile, the presented method is compared with existing methods such as MSE, composite MSE（CMSE） and refined composite MSE（RCMSE）. The results show that compared with MSE, CMSE and RCMSE, the proposed method enjoys significant advantages in terms of robustness, calculation efficiency and recognition accuracy, thereby providing a new idea for rotating machinery fault diagnosis based on entropy feature.