A noise-resistant fault diagnosis method based on adaptive feature mode decomposition and dual-stream context-aware network

Abstract In harsh industrial environments, extracting fault features from rolling bearing vibration signals is challenging due to intense background noise. To address the limitations of conventional networks—specifically their insufficient utilization of statistical characteristics in noisy signals and the tendency to overlook temporal correlations—this paper proposes an Anti-Noise Context-Aware Network (ANCAN). The framework integrates adaptive signal reconstruction with a hybrid deep learning architecture. Specifically, a novel metric, the Weighted Kurtosis-Energy Ratio-Envelope Spectrum Gini Index (WKEEGI), is constructed to adaptively guide Feature Mode Decomposition (FMD), enabling adaptive signal reconstruction that effectively highlights latent fault impulses. Furthermore, a dual-stream CNN-Transformer architecture integrated with a Preliminary Feature Enhancement (PFE) module is proposed. This framework synergistically captures local micro-structures and global temporal dependencies from multi-domain inputs. Experimental results on the HUST dataset demonstrate that ANCAN effectively captures long-sequence fault correlations, achieving a diagnostic accuracy of 97.28% under −6 dB strong noise, thereby validating its potential for practical industrial applications.