Multiwavelet denoising-guided dual-stream wavelet convolutional residual network for rolling bearing fault diagnosis

Abstract

Existing wavelet decomposition methods often fail to achieve a desirable trade-off between noise suppression and faithful preservation of weak fault signatures in high-noise environments, leading to the severe masking of incipient fault features. To address this limitation, a multiwavelet denoising-guided dual-stream wavelet convolutional residual network, termed MDWT-DSWIResNet, is proposed. First, to alleviate reconstruction distortion introduced by conventional denoising techniques, an orthogonal and symmetric SA4 multi-wavelet denoising scheme is developed. An entropy-driven adaptive decomposition strategy, together with an envelope kurtosis-guided thresholding function, is employed to enable adaptive component reconstruction, thereby producing bearing signals with enhanced signal-to-noise ratios. Second, a physically interpretable Morlet–Laplace dual-stream wavelet convolution module is constructed. By leveraging a dual-branch collaborative filtering mechanism, both periodic resonance components and transient impulsive features are simultaneously extracted, enabling accurate separation of multiple fault-related signatures in bearing vibration signals. Finally, to overcome the limitations of conventional classifiers in long-sequence modeling, including restricted receptive fields and inadequate capture of long-range temporal dependencies, a large-kernel attention-based residual block is incorporated to enhance global temporal representation capability for fault pattern classification. Extensive experiments conducted on the CWRU, Paderborn, and in-house laboratory datasets demonstrate that the proposed method achieves superior diagnostic accuracy under extreme noise, mixed operating conditions, and compound fault scenarios.