A Two-Stage Manifold Learning Framework for Machinery Fault Diagnosis

This study presents a new fault diagnosis method based on a two-stage manifold learning framework to further improve fault diagnosis accuracy. First of all, nonlinear de-noising method with unsupervised manifold learning is presented, by combining advantages of manifold learning in mining of nonlinear structure and phase space reconstruction in representation of signal and noise spatial distribution. Then, the frequency spectrum of vibration signals after de-noising is used for fault feature extraction. In order to reduce the high dimensionality and remove redundant information of frequency spectrum, an improved supervised local tangent space alignment (ISLTSA) is proposed. ISLTSA further increases interclass distance and further reduces intraclass distance, and as a result the extracted fault features are more identifiable. At last, the extracted low-dimensional fault features are inputted into a pattern recognition method for fault identification. A fault diagnosis case in bearings is studied to verify the effectiveness of the proposed method.