Scholarlink

Purpose <p>Accurate bearing fault diagnosis under strong noise interference remains a significant challenge in condition monitoring. This study proposes an integrated framework for robust bearing fault diagnosis under noisy conditions.</p> Methods <p>A Dynamic Time Warping (DTW)-based Intrinsic Mode Function (IMF) qualification metric is used to suppress noise-dominant components. Multiresolution time–frequency characterization is achieved by combining the Continuous Wavelet Transform (CWT) and the two-dimensional Discrete Wavelet Transform (2D-DWT), along with a newly designed set of wavelet-based features. A Tree Seed Algorithm (TSA)-based feature space refinement strategy is then applied to select highly informative features, which are subsequently classified using a Bidirectional Long Short-Term Memory (BiLSTM) network.</p> Results <p>The proposed framework is validated through three case studies conducted under varying noise levels. The method achieved classification accuracies of 91.13% on the CWRU dataset, 78.04% on the JNU dataset, and 99.4% on the MFPT dataset, even under − 12 dB of white noise.</p> Conclusion <p>Results across multiple datasets demonstrate the effectiveness of the proposed methodology, making it suitable for real-world industrial condition-monitoring applications under noisy operating conditions.</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

Data-Driven Noise Suppression and Multiresolution Feature Fusion with Feature Space Refinement for Bearing Fault Diagnosis

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Andrews Athisayam,
  • Abisha D.

摘要

Purpose

Accurate bearing fault diagnosis under strong noise interference remains a significant challenge in condition monitoring. This study proposes an integrated framework for robust bearing fault diagnosis under noisy conditions.

Methods

A Dynamic Time Warping (DTW)-based Intrinsic Mode Function (IMF) qualification metric is used to suppress noise-dominant components. Multiresolution time–frequency characterization is achieved by combining the Continuous Wavelet Transform (CWT) and the two-dimensional Discrete Wavelet Transform (2D-DWT), along with a newly designed set of wavelet-based features. A Tree Seed Algorithm (TSA)-based feature space refinement strategy is then applied to select highly informative features, which are subsequently classified using a Bidirectional Long Short-Term Memory (BiLSTM) network.

Results

The proposed framework is validated through three case studies conducted under varying noise levels. The method achieved classification accuracies of 91.13% on the CWRU dataset, 78.04% on the JNU dataset, and 99.4% on the MFPT dataset, even under − 12 dB of white noise.

Conclusion

Results across multiple datasets demonstrate the effectiveness of the proposed methodology, making it suitable for real-world industrial condition-monitoring applications under noisy operating conditions.