Abstract
Rolling bearings are crucial in rotating machinery, and combining natural and characteristic frequencies improves fault detection. However, natural frequencies face challenges like feature extraction difficulties and drift, necessitating resonance peak information supplementation. Existing methods for extracting resonance peaks often struggle with low quality, false peaks, and merging issues. This paper introduces a novel resonance peak extraction method based on auditory saliency (RESAS), inspired by the human auditory system. RESAS combines Gammatone filtering, multi-scale Gaussian filtering, and lateral inhibition to simulate auditory attention and efficiently extract resonance peaks. A resonance peak saliency map (RPSP) is generated, from which features are extracted and used as input to an improved random forest model (TF-RF) for fault classification. Tests on the QPZZ-II Fault Simulation Test Bench and KWCU data show that the method effectively identifies bearing faults at various speeds and loads, demonstrating its strong potential for application. Furthermore, due to its broadband characteristics and capacity to excite the system's natural frequencies, this method has potential for scalability to other impact-type fault systems.