A fault diagnosis model based on composite multi-scale bubble entropy was proposed to address the issues of insufficient diagnostic effectiveness and poor generalization of conventional models when dealing with limited fault samples of rolling bearings, multiple damage categories, and uneven distribution of fault samples. Initially, the original signals were decomposed into multiple product components using robust local mean decomposition, Based on the Pearson correlation coefficient, significant components were selected and reconstructed. Subsequently, composite theory and multi-scale theory were incorporated into bubble entropy to develop composite multi-scale bubble entropy, thereby enhancing the feature extraction capability for the reconstructed signals. To improve the recognition performance of the back propagation neural network, the black-winged kite algorithm was employed to optimize its parameters. Through experimental applications on three different bearing datasets, it was found that the model achieved an average diagnostic accuracy higher than 98% under small-sample conditions, with a sample standard deviation of 0.36, demonstrating high diagnostic accuracy and stability. Result provides a novel approach for rolling bearing fault diagnosis with limited samples.