Blind spectrum-guided adaptive cyclic refinement network for bearing fault diagnosis under complex interference

Abstract

The extraction of weak bearing fault features remains a formidable challenge in the field of mechanical health monitoring, particularly under strong background noise and non-periodic random shocks. Although blind deconvolution methods can effectively eliminate transmission path effects, their reliability is severely restricted by an inherent reliance on prior knowledge and the sparsity trap induced by large-amplitude random shocks under low signal-to-noise ratio conditions. To address these limitations, this paper proposes a novel unsupervised feature extraction method designated as the spectrum-guided adaptive cyclic refinement network (SG-ACR). In this paper, a global blind spectral anchor capture mechanism is designed to estimate the fault characteristic frequency by integrating multi-band fusion and the harmonic product spectrum, thereby resolving the initial parameterization problem. Subsequently, a learnable convolutional layer featuring the harmonic mean cyclic attention (HMCA) mechanism is constructed. The HMCA dynamically suppresses random shocks and guides the network to converge accurately to the genuine periodic fault components by enforcing strict temporal synchronization constraints. Additionally, a physics-informed Gaussian shaping strategy is introduced to transform data-driven weights into interpretable resonance parameters, thereby mitigating in-band noise overfitting. The superior robustness of the SG-ACR is demonstrated through comprehensive simulations and experimental verifications. Using a practical bearing dataset, the SG-ACR successfully extracts the fundamental fault frequency and its first five harmonics with minimal background noise, significantly outperforming traditional methods.