Despite the recent advances on intelligent fault diagnosis of rolling element bearings, existing researches mostly assume training and testing data are drawn from the same distribution. However, due to variation of operating condition, environmental noise and so forth, domain shift phenomenon generally exists, that significantly results in diagnosis performance deterioration. To address cross-domain problems, latest researches preferably apply domain adaptation techniques on marginal data distributions. However, it is usually assumed sufficient testing data are available for training, that is not in accordance with most transfer tasks in real industries where only data in machine healthy condition can be collected in advance. This paper proposes a novel cross-domain fault diagnosis method based on deep generative neural networks. By artificially generating fake samples for domain adaptation, the proposed method is able to provide reliable diagnosis results when the testing data in different machine fault conditions are not available for training. The experimental results suggest the proposed method offers a promising approach for industrial applications.