Multiple-domain manifold for feature extraction in machinery fault diagnosis

• Phase space reconstruction is utilized to construct 2-D matrices representing signals in time and frequency domains.
• Singular value decomposition is employed to calculate the singular values of multiple domains as preliminary features.
• Manifold learning is introduced to revise the singular values to obtain the MDM features for fault diagnosis.
• Practical engineering cases verified the advantages of MDM features in machinery fault diagnosis.

Extracting features of vibration signals is a key technology in mechanical device state monitoring and fault diagnosis. This study proposes a novel multiple-domain manifold (MDM) method to achieve representative features based on singular value decomposition (SVD) and manifold learning. MDM features are generated by three main steps: first, phase space reconstruction is applied to signals in time domain and frequency domain to achieve a reconstructed 2-D space, respectively; second, SVD is used to calculate singular values (SVs) in the reconstructed spaces, as well as the improved Hilbert–Huang spectrum of the signal; and finally, manifold learning is employed to extract the MDM features by revising the SVs. The MDM features can reveal the intrinsic information of signals, and exhibit high-stability denoising effect. Moreover, the low-dimension property is beneficial for an effective diagnosis. The validity of MDM is confirmed by detecting bearing and gear faults, which demonstrates the advantages and potential practical applications of the method for recognizing mechanical fault patterns.