Industrial faults are rare and sporadic， so the number of fault database samples is generally insufficient. The condition of small sample fault database can easily cause problems such as over-fitting in traditional deep learning， which affects the accuracy of diagnosis. In order to increase the sample size， obtain the fault information， and improve the accuracy of fault diagnosis under the condition of small sample fault database， a method based on deep convolutional generative adversarial networks（DCGAN） is proposed. The checkerboard problem of traditional algorithm is improved through combining the traditional DCGAN algorithm with the nearest neighbor interpolation up-sampling and convolution（USCONV layer）. After feature extraction and training of three-channel wavelet images， the model output realistically generated images. The model can accurately expand and enrich the sample set under the condition of small samples fault database， alleviate over fitting and other problems， and improve accuracy of diagnosis. The results show that the USCONV layer can significantly improve the checkerboard problem. In addition， the test accuracy of the diagnosis model for the test set containing various fault conditions before and after the expansion of small-sample-database is increased from 91.67% to 98.96%， which demonstrate the method is feasible and effective.