Scholarlink

<p>Bearing fault diagnosis is critical for the predictive maintenance of industrial equipment. However, existing deep learning methods frequently encounter bottlenecks, such as feature extraction degradation and insufficient generalization capability, when operating under complex working conditions and severe noise interference. To address these issues, this paper proposes a&#xa0;novel dual-branch collaborative fault diagnosis network, termed C‑SwinNet, which integrates a&#xa0;Convolutional Neural Network (CNN) and a&#xa0;Swin Transformer. In the data preprocessing stage, a&#xa0;multi-modal RGB image generation strategy is proposed. By fusing the Continuous Wavelet Transform (CWT), Short-Time Fourier Transform (STFT), and Recurrence Plot (RP), this strategy deeply aligns the physical attributes of time-frequency signals with the channel perception mechanisms of deep vision networks, thereby providing the model with highly efficient, feature-complementary inputs. To overcome the technical challenges associated with architectural fusion and severe noise, C‑SwinNet deeply integrates three core mechanisms. First, to mitigate strong noise interference, the Swin Transformer branch incorporates a&#xa0;Multi-Dimensional Spatial Channel Attention (MDSCA) module, which effectively filters local high-frequency noise and enhances the robust capturing capability of global information. Second, a&#xa0;Fusion Block is designed to achieve rigorous spatial alignment of cross-modal features via pyramid pooling and cross-attention mechanisms, successfully bridging the semantic gap between local details and global contexts. Third, a&#xa0;dynamic routing module, DynamicGLU, is introduced at the terminal stage of the network. Utilizing a&#xa0;three-weight gating mechanism, it adaptively adjusts the collaborative proportion of the dual-branch features based on real-time operating conditions. Comprehensive experiments conducted on five datasets, including CWRU, MFPT, JNU, Ottawa, and a&#xa0;laboratory self-built dataset, demonstrate that C‑SwinNet achieves a&#xa0;macro-averaged accuracy of 98.7 ± 0.5% and a&#xa0;macro-F1 score of 0.986 ± 0.006. Notably, under Gaussian white noise and pink noise environments, its average accuracy outperforms state-of-the-art models, such as Diagnosisformer and Swin-FFRN, by margins ranging from 2% to 6%. Supported by extensive ablation studies and complexity analyses, these results thoroughly validate the remarkable advantages of the proposed model in terms of diagnostic accuracy, anti-noise robustness, and computational efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

C-SwinNet: hybrid CNN-Swin transformer integration for enhanced fault diagnosis of rolling bearings under noisy conditions

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Zhihan Kang,
  • Xiaofeng Yue,
  • Yuan Zhou,
  • Chengxiao Liu

摘要

Bearing fault diagnosis is critical for the predictive maintenance of industrial equipment. However, existing deep learning methods frequently encounter bottlenecks, such as feature extraction degradation and insufficient generalization capability, when operating under complex working conditions and severe noise interference. To address these issues, this paper proposes a novel dual-branch collaborative fault diagnosis network, termed C‑SwinNet, which integrates a Convolutional Neural Network (CNN) and a Swin Transformer. In the data preprocessing stage, a multi-modal RGB image generation strategy is proposed. By fusing the Continuous Wavelet Transform (CWT), Short-Time Fourier Transform (STFT), and Recurrence Plot (RP), this strategy deeply aligns the physical attributes of time-frequency signals with the channel perception mechanisms of deep vision networks, thereby providing the model with highly efficient, feature-complementary inputs. To overcome the technical challenges associated with architectural fusion and severe noise, C‑SwinNet deeply integrates three core mechanisms. First, to mitigate strong noise interference, the Swin Transformer branch incorporates a Multi-Dimensional Spatial Channel Attention (MDSCA) module, which effectively filters local high-frequency noise and enhances the robust capturing capability of global information. Second, a Fusion Block is designed to achieve rigorous spatial alignment of cross-modal features via pyramid pooling and cross-attention mechanisms, successfully bridging the semantic gap between local details and global contexts. Third, a dynamic routing module, DynamicGLU, is introduced at the terminal stage of the network. Utilizing a three-weight gating mechanism, it adaptively adjusts the collaborative proportion of the dual-branch features based on real-time operating conditions. Comprehensive experiments conducted on five datasets, including CWRU, MFPT, JNU, Ottawa, and a laboratory self-built dataset, demonstrate that C‑SwinNet achieves a macro-averaged accuracy of 98.7 ± 0.5% and a macro-F1 score of 0.986 ± 0.006. Notably, under Gaussian white noise and pink noise environments, its average accuracy outperforms state-of-the-art models, such as Diagnosisformer and Swin-FFRN, by margins ranging from 2% to 6%. Supported by extensive ablation studies and complexity analyses, these results thoroughly validate the remarkable advantages of the proposed model in terms of diagnostic accuracy, anti-noise robustness, and computational efficiency.