Adaptive Multi-scale Boosting Dictionary Learning for Bearing Fault Diagnosis

Extracting the fault impulses caused by localized faults is crucial for detecting bearing faults. However, the extraction task is challenging since the impulse signal is easily submerged by strong noise and complex harmonic interference. In this paper, an adaptive multi-scale boosting dictionary learning (AMBDL) method is proposed, which can effectively extract weak fault signals even in early fault stage, so that the bearing fault can be diagnosed timely. Specifically, a multi-scale boosting dictionary learning (MBDL) model is first constructed, which integrates the strengthen-operate-subtract (SOS) boosting strategy and dictionary learning into multi-scale transform to iteratively boost the learning ability of multi-scale dictionary. Second, a robust fault-sensitive index, adaptive periodic modulation intensity (APMI), is designed for subband screening to remove subbands that mainly contain interference, enabling MBDL to effectively focus on enhancing and learning fault features in the optimal subbands. Third, a threshold estimation method is constructed to adaptively set the threshold parameters in sparse coding stage of the MBDL, which is suitable for real-time fault diagnosis. The analysis and comparison results of simulation and bearing failure experiments show that AMBDL is superior to some advanced methods in fault impulse extraction, while requiring lower computational costs and achieving adaptivity.