Multichannel Sparse Bayesian Learning With GAMP Integration for Bearing Fault Diagnosis

Data collected from multiple sensors contains rich fault information for bearing fault diagnosis. However, it is difficult to leverage these joint information due to the following reasons: (1) signals from different channels often have asynchronous sampling lengths, which can greatly hinder subsequent processing; (2) unpaired frequency transformations can lead to significant algorithmic degradation; and (3) handling multi-channel data simultaneously can be quite time-consuming. To solve above three challenges, in this paper, we newly propose a multi-channel sparse Bayesian learning (SBL) based approach for bearing fault diagnosis. The key innovations are: (1) extend the existing sparse frequency learning model into multi-channel form, accommodating variable-length signals across different channels; (2) identify the fault frequencies from the signal envelopes without any frequency transformations by leveraging the common sparsity of the fault frequencies-of-interest in the frequency domain; and (3) construct a real-valued data model with a special unitary matrix, and integrate the generalized approximate message passing (GAMP) method into the Bayesian framework to enhance the fault detection performance while reducing computational complexity. Since the message passing has the property of element-wise operation and can provide exact marginal posteriors, the proposed method can achieve state-of-the-art performance with low-complexity. Both simulation and real datasets can demonstrate it.