Using deep learning to enhance dataset has become a hot topic in various fields. That is，using limited dataset to generate more simulated dataset. Different from the current mainstream generative adversarial network（GAN）and its variant algorithms，based on the idea of sample resolution enhancement，this paper proposes a simple and effective algorithm—an efficient sub-pixel fully connected neural network（ESPFCN）. The principle of ESPFCN is：it performs fully-connected operation on the raw input samples and the results of four-channel multi-feature maps are output；through the fully connected sub-pixel layer，the low resolution features of the four channels are arranged periodically，and a set of high resolution features is obtained，which enhances the sample resolution. Finally，a set of special bearing experiment is set up to evaluate the performance of the generated model. The experimental results verify the effectiveness of the ESPFCN framework，and the feature learning process of ESPFCN is visualized.