An adaptive orthogonality-constrained robust dictionary learning approach and its application to bearing fault diagnosis

Extracting weak fault features from vibration signals contaminated with strong noise, especially non-Gaussian noise, is a challenging task for early fault diagnosis of bearings. In this article, a novel adaptive orthogonality-constrained robust dictionary learning (AOCRDL) method for bearing fault diagnosis is developed to address this issue. First, an $\alpha $ -divergence-based robust dictionary learning (RDL) approach was introduced to enhance the capability of sparse representation techniques in handling non-Gaussian noise. Second, orthogonalization process of atoms is embedded into the dictionary learning process to obtain an optimized dictionary for recovering impulse sequences related to faults. Third, the adaptive determination method of key parameters in the AOCRDL was developed to ensure optimal performance of the approach. Simulations and experimental results demonstrate that AOCRDL method can effectively extract early bearing fault features, especially in scenarios with strong non-Gaussian noise.