Fast Frequency Sparsity Learning Approach for Missing Data-Resistant Bearing Fault Diagnosis

Bearing fault diagnosis in the presence of missing data poses a significant challenge due to the substantial information loss brought by incomplete databases. Current state-of-the-art techniques for addressing missing data, whether involving data imputation or listwise deletion, are susceptible to notable algorithmic deterioration as the proportion of missing samples within a database increases. In this article, we novelly propose a missing data-resistant Bayesian optimal approach to directly learn the fault frequencies from the incomplete signal envelope. We first introduce an enhanced sparse frequency learning model to avoid time-domain signal recovery and frequency transformation. Then, we characterize the missing samples within the Bayesian framework to pave the way to extract fault frequencies of interest without compromising performance under missing data conditions. Finally, we resort to the generalized approximate message-passing (GAMP) method to marginalize the missing data automatically and identify the fault frequencies accurately. The proposed method offers a considerable performance improvement while maintaining low computational complexity. Both simulation and real dataset results verify the superiority of the proposed method.