Scholarlink

<p>The Transformer network has powerful parallelization and global information capturing capabilities by utilizing the self-attention mechanism, which can effectively capture long-term dependencies between sequence elements. Compared to traditional deep learning methods, Transformer models global contextual information without recursive or convolutional operations, and exhibits greater adaptability when dealing with non-stationary signals. In the field of fault diagnosis, this architecture is particularly good at capturing fault features in vibration signals across time scales, providing a new solution for fault pattern recognition under complex operating conditions. However, the performance is easily affected under small sample conditions or in the presence of strong non-stationary noise. The data samples collected in engineering practice often come from different working conditions and contain complex noise, which further increases the challenge of model training. Diagnostic methods based on one-dimensional vibration data are difficult to intuitively mine the local features of fault signals, which limits the ability of deep-level feature extraction. To address the above problems, this paper proposes a Wavelet-Driven Transformer network-based Dual-scale intelligent feature screening Fault Detector (WDT-DFFD). Firstly, an adaptive time-frequency analysis layer is built using continuous wavelet convolutional layers, and the scaling and translation parameters are learned directly from the original data to convert the time-domain signal into time-frequency domain images, which enhances the feature extraction and signal denoising capabilities. Secondly, the Transformer in Transformer (TNT) network is utilized to build a Dual-scale Fault Detector, and the Bayesian variational attention mechanism is introduced as an intelligent feature filter, which enhances the network's ability to capture image details and fault information. Finally, experiments on the bearing vibration dataset and gearbox vibration dataset show that the method maintains high diagnostic accuracy and robustness under small samples and strong noise 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

Wavelet-Driven Transformer-Based Dual-Scale Intelligent Feature Selection Fault Detector

  • 期刊论文

  • 发表日期：

    2025年7月7日

查看全文
  • Zhiwu Shang,
  • Xinqiang Mao,
  • Leyi Yao,
  • Hongchuan Cheng

摘要

The Transformer network has powerful parallelization and global information capturing capabilities by utilizing the self-attention mechanism, which can effectively capture long-term dependencies between sequence elements. Compared to traditional deep learning methods, Transformer models global contextual information without recursive or convolutional operations, and exhibits greater adaptability when dealing with non-stationary signals. In the field of fault diagnosis, this architecture is particularly good at capturing fault features in vibration signals across time scales, providing a new solution for fault pattern recognition under complex operating conditions. However, the performance is easily affected under small sample conditions or in the presence of strong non-stationary noise. The data samples collected in engineering practice often come from different working conditions and contain complex noise, which further increases the challenge of model training. Diagnostic methods based on one-dimensional vibration data are difficult to intuitively mine the local features of fault signals, which limits the ability of deep-level feature extraction. To address the above problems, this paper proposes a Wavelet-Driven Transformer network-based Dual-scale intelligent feature screening Fault Detector (WDT-DFFD). Firstly, an adaptive time-frequency analysis layer is built using continuous wavelet convolutional layers, and the scaling and translation parameters are learned directly from the original data to convert the time-domain signal into time-frequency domain images, which enhances the feature extraction and signal denoising capabilities. Secondly, the Transformer in Transformer (TNT) network is utilized to build a Dual-scale Fault Detector, and the Bayesian variational attention mechanism is introduced as an intelligent feature filter, which enhances the network's ability to capture image details and fault information. Finally, experiments on the bearing vibration dataset and gearbox vibration dataset show that the method maintains high diagnostic accuracy and robustness under small samples and strong noise conditions.