Group-Sparsity Learning Approach for Bearing Fault Diagnosis

Fault impulse extraction under strong background noise and/or multiple interferences is a challenging task for bearing fault diagnosis. Sparse representation has been widely applied to extract fault impulses and can achieve state-of-the-art performance. However, most of the current methods rely on carefully tuning several hyper-parameters and suffer from possible algorithmic degradation due to the approximate regularization and/or heuristic sparsity model. To overcome these drawbacks, we present a sparse Bayesian learning (SBL) framework for bearing fault diagnosis, and then propose two group-sparsity learning algorithms to extract fault impulses, where the first one exploits the group-sparsity of fault impulses only; while the second one utilizes additional periodicity behavior of fault impulses. Due to the inherent learning capability of the SBL framework, the proposed algorithms can tune hyper-parameters automatically and do not require any prior knowledge. Another advantage is that our solutions are maximum a posteriori estimators in the sense of Bayesian optimality, which can yield higher accuracy. Results on both simulated and real datasets demonstrate the superiority of the developed algorithms.