Scholarlink

<p>In practical industrial environments, the fault signals of rolling bearings are often affected by noise interference and limited sample availability. To address the challenge of fault diagnosis for rolling bearings under noisy and small-sample conditions, this study proposes an improved one-dimensional convolutional neural network (1D-CNN) method based on frequency-domain features, referred to as Gated Multi-Head Attention 1D-CNN (GMA-1DCNN). First, the raw vibration signals are transformed into the frequency domain using the Fast Fourier Transform (FFT), which enhances spectrum stability and suppresses noise interference. Then, a wide convolutional kernel is introduced in the early network layers to enlarge the receptive field and strengthen the modeling of long-range dependencies. Furthermore, a gated multi-head attention (GMA) module is designed to adaptively allocate attention weights and thereby highlight critical feature representations. Experimental results demonstrate that the proposed method achieves average diagnostic accuracies of 97.93% and 99.14% on two datasets when only five samples per class are available. Under a sample size of 30 and a 0 dB noise condition, the accuracies remain as high as 96.96% and 98.92%, respectively. These results verify the robustness and strong generalization capability of the method in both small-sample and noisy environments.</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

Frequency-domain small-sample bearing fault diagnosis with Wide-Kernel convolution and gated attention

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Yinghui Wang,
  • Xijun Zhang,
  • Jiwen Li,
  • Xiaogang Liu

摘要

In practical industrial environments, the fault signals of rolling bearings are often affected by noise interference and limited sample availability. To address the challenge of fault diagnosis for rolling bearings under noisy and small-sample conditions, this study proposes an improved one-dimensional convolutional neural network (1D-CNN) method based on frequency-domain features, referred to as Gated Multi-Head Attention 1D-CNN (GMA-1DCNN). First, the raw vibration signals are transformed into the frequency domain using the Fast Fourier Transform (FFT), which enhances spectrum stability and suppresses noise interference. Then, a wide convolutional kernel is introduced in the early network layers to enlarge the receptive field and strengthen the modeling of long-range dependencies. Furthermore, a gated multi-head attention (GMA) module is designed to adaptively allocate attention weights and thereby highlight critical feature representations. Experimental results demonstrate that the proposed method achieves average diagnostic accuracies of 97.93% and 99.14% on two datasets when only five samples per class are available. Under a sample size of 30 and a 0 dB noise condition, the accuracies remain as high as 96.96% and 98.92%, respectively. These results verify the robustness and strong generalization capability of the method in both small-sample and noisy environments.