<p>Aiming at the problems of information loss, feature redundancy and unsatisfactory diagnosis accuracy when using traditional multiscale entropy methods and feature reduction methods to diagnose rotating machinery faults, a feature extraction method based on a multiscale entropy fusion strategy and a GA-RL-LDA model is proposed in this paper. Firstly, the multiscale fluctuation dispersion entropy (MFDE), the refined composite multiscale dispersion entropy (RCMDE) and the refined composite multiscale fluctuation dispersion entropy (RCMFDE) of the collected vibration signal are calculated to form an original feature set. Then, based on the ReliefF algorithm and Laplacian score (LS), an RL index is constructed for feature sensitivity evaluation. After that, combing the RL with Linear discriminant analysis (LDA) and using genetic algorithm (GA) to optimize the uncertain parameters, a GA-RL-LDA model is proposed for feature reduction. Finally, the reduced feature subset is input into support vector machine (SVM) for fault classification. The experiment utilized data from Unit 3 of the SK Hydropower Station and bearing data from Case Western Reserve University, achieving diagnostic accuracies of 95.2381% and 97.3333%, respectively. In the 105 test samples from Unit 3 of the SK Hydropower Station, only 5 samples were misclassified, while in the 150 test samples from Case Western Reserve University, only 4 samples were misclassified. Compared with different information entropy and optimization strategies, the results show that the proposed method can more effectively extract fault sensitive features and accurately diagnose rotating machinery faults even with a small number of training samples.</p>