Scholarlink

<p>To address the challenge of rotating machinery fault features being easily overwhelmed by noise in harsh environments, this paper proposes a dual-branch parallel diagnosis method. The method integrates a multi-scale dynamic convolutional attention mechanism with an adaptive sparse attention transformer encoder. Firstly, short-time fourier transform (STFT) is employed to convert raw vibration signals into two-dimensional time-frequency grayscale images, preserving both temporal and spectral information of fault features. Secondly, in the convolutional neural network (CNN) branch, a multi-scale feature extraction module and an enhanced omni-dimensional dynamic convolutional attention block are designed to adaptively modulate and enhance feature representations. Depthwise separable convolutions are further incorporated to improve feature extraction efficiency. In the transformer encoder branch, an adaptive sparse attention mechanism is introduced, thereby enhancing robustness under low signal-to-noise ratio (SNR) conditions. Finally, experimental results on three different datasets demonstrate that the proposed method maintains high diagnostic accuracy even under extreme noise levels ranging from -6 to -12 decibels (dB), significantly outperforming lightweight models and traditional residual networks, thereby validating its superior noise robustness under complex working 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

A Dual-Branch Dynamic Attention and Sparse Transformer Network for Noise-Robust Bearing Fault Diagnosis

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Shuo Mu,
  • DeChen Yao,
  • JianWei Yang,
  • Bin Zhu

摘要

To address the challenge of rotating machinery fault features being easily overwhelmed by noise in harsh environments, this paper proposes a dual-branch parallel diagnosis method. The method integrates a multi-scale dynamic convolutional attention mechanism with an adaptive sparse attention transformer encoder. Firstly, short-time fourier transform (STFT) is employed to convert raw vibration signals into two-dimensional time-frequency grayscale images, preserving both temporal and spectral information of fault features. Secondly, in the convolutional neural network (CNN) branch, a multi-scale feature extraction module and an enhanced omni-dimensional dynamic convolutional attention block are designed to adaptively modulate and enhance feature representations. Depthwise separable convolutions are further incorporated to improve feature extraction efficiency. In the transformer encoder branch, an adaptive sparse attention mechanism is introduced, thereby enhancing robustness under low signal-to-noise ratio (SNR) conditions. Finally, experimental results on three different datasets demonstrate that the proposed method maintains high diagnostic accuracy even under extreme noise levels ranging from -6 to -12 decibels (dB), significantly outperforming lightweight models and traditional residual networks, thereby validating its superior noise robustness under complex working conditions.