EWT-CEC-RGATv2: An enhanced collaborative correntropy driven and interpretable method for gear signal denoising

Vibration signal denoising is crucial for reliable mechanical health monitoring; however, environmental noise often obscures informative features and degrades diagnostic accuracy. Existing modal decomposition and graph-based methods usually neglect global inter-modal dependencies and lack interpretability in noise representation learning. To address these limitations, this article proposes an interpretable synergistic denoising framework termed EWT-CEC-RGATv2, which integrates empirical wavelet transform (EWT), a comprehensive enhancement correntropy (CEC) module, and a residual graph attention network (RGATv2). EWT first decomposes the vibration signal into adaptive modal components that are organized as graph nodes, while edges represent correntropy-based relationships between modes. The proposed CEC module introduces an information-theoretic correntropy synergy mechanism to enhance and dynamically weight inter-modal dependencies, enabling collaborative noise suppression across modalities. Based on the enhanced relational representation, RGATv2 employs residual connections and multihead attention to learn noise-to-signal mappings while preserving intrinsic signal structures. Experimental results on both simulated and planetary gearbox datasets demonstrate that the proposed framework consistently achieves superior denoising performance across a broad noise range from −10 to +10 dB, outperforming state-of-the-art methods in terms of robust signal-to-noise ratio, root mean square error, and normalized correlation coefficient. Furthermore, visualization of graph structures and attention weights provides transparent and interpretable insights into the denoising process.