To deal with the problem that features learned by a traditional auto-encoder(AE) are less discriminative due to unsupervised manner, we propose a cluster compact auto-encoder(CCAE). First of all, a fuzzy C-means algorithm is used to cluster samples to get pseudo labels, where the optimal number of clusters is determined by the PBMF index. Then, a cluster compact regularization(CCR) is established based on the pseudo labels, which embeds discriminant information indicating categories of samples. Finally, the CCR is combined with the AE to constitute the CCAE's loss function. Discriminant ability of the proposed method can be enhanced via the pseudo labels that incorporate discriminant information indicating categories, so as to improve the diagnostic performance greatly. The effectiveness of the proposed method is verified on rotating machinery gear and bearing datasets. The proposed CCAE can be widely applicable to the feature extraction stage of rotating machinery fault diagnosis, which provides a solution for engineers to realize automatic extraction of discriminative features.