Nowadays, the measuring and monitoring data collected from rotating machinery exists a high-dimensional and massive status in quo, but the acquisition of the valuable labeled information is difficult. To address this challenge, an intelligent fault diagnosis method of rotating machinery based on semisupervised deep sparse auto-encoder (SSDSAE) with local and nonlocal information is presented in this article. Primarily, the collected vibration spectrum signals are fed into the constructed SSDSAE algorithm to implement fault feature extraction, and after that the extracted sparse discriminant features are fed into the back-propagation (BP) classifier for fault diagnosis. Accordingly, the unlabeled samples in SSDSAE are used to define local and nonlocal structural information for the original DSAE, and then the topological regularization terms are obtained by minimizing intraclass compactness (locality) and maximizing interclass separability (nonlocality). Moreover, the weighted cross-entropy (WCE) techniques in SSDSAE are used to improve the generalization performance of the fault diagnosis model for labeled data. Eventually, the availability of the proposed method is validated with data from experimental bearing fault and industrial rotor fault, respectively.