As a breakthrough in the field of machine fault diagnosis, deep learning has great potential to extract more abstract and discriminative features automatically without much prior knowledge compared with other methods, such as the signal processing and analysis-based methods and machine learning methods with shallow architectures. One of the most important aspects in measuring the extracted features is whether they can explore more information of the inputs and avoid redundancy to be representative. Thus, a stacked sparse autoencoder (SAE)-based machine fault diagnosis method is proposed in this paper. The penalty term of the SAE can help mine essential information and avoid redundancy. To help the constructed diagnosis network further mine more abstract and representative high-level features, the collected non-stationary and transient signals are preprocessed with ensemble empirical mode decomposition and autoregressive (AR) models to obtain AR parameters, which are extracted based on the intrinsic mode functions (IMFs) and regarded as the low-level features for the inputs of the proposed diagnosis network. Only the first four IMFs are considered, because fault information is mainly reflected in high-frequency IMFs. Experiments and comparisons are complemented to validate the superiority of the presented diagnosis network. Results fully demonstrate that the stacked SAE-based diagnosis method can extract more discriminative high-level features and has a better performance in rotating machinery fault diagnosis compared with the traditional machine learning methods with shallow architectures.