A clustering low-rank approach for aero-enging bearing fault detection

The highly overlapping distortion characteristic of high speed aero-engine bearing faults violates the fundamental assumption of popular bearing fault diagnostic techniques which assume that every impulse has a distinct exponential-decaying pattern. Therefore, a tailored clustering low rank framework (coined as CluLR) is proposed for the feature detection of aero-engine bearings. This work firstly explores the underlying prior information that fault features demonstrate multiple similarity structures in a transformed data matrix obtained through employing an elaborately designed partition operator. Then, incorporating the clustering procedure into low-rank regularization model, the proposed CluLR guarantees that different similarity information is reliably concentrated onto their matched low-rank domains, which effectively eliminates the singular value overlapping coherent pathology. Consequently, weak features as well as strong features could be detected simultaneously. Moreover, an alternative minimization algorithm adopted from block coordinate descent framework is developed to solve the two-stage nonsmooth and nonconvex problem. Lastly, compared with the state-of-the-art bearing diagnosis techniques, the proposed CluLR's superiority is sufficiently verified through its application to the experimental data from an aero-engine bearing under 25000 rev/min for overlapping distorted feature detection tasks.