Low-rank enhanced convolutional sparse feature detection for accurate diagnosis of gearbox faults

Abstract

It is a challenge problem to accurately recognize damage distribution pattern for multi-stage industrial gearboxes in filed, due to entangled relationships between strong interferences/noises and complicate transfer path modulations. In this work, a tailored two-stage strategy (LR-CSL) based on low-rank representation and convolutional sparse learning is proposed. Based on the periodic similarity of focused features, a weighted low-rank stage is firstly utilized to suppress strong interferences and noises, which provides a cornerstone to enhance blind deconvolution methods. Then, a convolutional sparse stage is adopted to mitigate the transfer path modulation by enforcing one nonnegative bounded regularizer, which guarantees the reliable recovery of impulsive source envelopes. Lastly, the damage distribution patterns could be reliably confirmed by directly referring to the recovered source envelopes (rather than modulated waveforms) and gearbox dynamics. Comprehensive health evaluations to one 750 kW wind turbine drivetrain are performed blindly and gear surfaces with multiple weak spalling patterns are recognized accurately. Moreover, the spalling fault evolution process is deduced and maintenance guidances are allocated. Further analysis also confirms the first low-rank stage plays a necessary and important role in boosting LR-CSL’s deconvolution capability. Lastly, quantitative evaluations demonstrate that our LR-CSL method achieves a higher diagnostic accuracy than state-of-the-art fault diagnosis techniques.