In this paper, a new sparsity and neighborhood preserving-deep extreme learning machine (SNPDELM) based on sparsity and neighbor preserving theory is proposed for fault diagnosis of rolling bearings. Firstly, a structure of extreme learning machine-autoencoder (ELM-AE) is constructed by combining the extreme learning machine with the auto-encoder, and the hidden layer of the extreme learning machine is layered by auto-encoder. Then, the theories of sparse representation and neighbor representation are introduced to the deep network. During the projection process, the global structure of the data is maintained by sparse representation and the local manifold structure of the data is maintained by the neighbor representation. The deep features of the data are extracted without supervision successively. Finally, the data are classified by solving least squares. This method is applied to diagnose faults of the fan rolling bearings. Compared with other algorithms such as deep extreme learning machine (DELM), extreme learning machine (ELM), stacked autoencoder (SAE), and convolution neural network (CNN), the experimental results show that the algorithm proposed in this paper has higher accuracy and stability.