This paper is concerned with the mechine fault diagnosis problem of nonlinearities, and diversities and complexities of fault symptoms. Based on the wavelet packet sample entropy and manifold learning, a model of fault feature extraction is proposed. Firstly, to extract the initial rolling bearing fault feature, the model calculates the sample entropy of the signal reconstructed by using the wavelet packet decomposition and reconstruction method. Then the Local Tangent Space Alignment (LTSA) for further extraction is applied. In this sense, the model greatly reduces the complexity of feature data. In the meanwhile, the structure information in the whole geometry of the fault signal can be reserved. Moreover, the proposed model also enhances the classification performance of the entire fault mode
identification. Finally, the support vector machine is used to classify the feature extraction from the proposed model. The primary feature extraction and further feature extraction of classification results are then compared to validate the superiority of the proposed model.