Scholarlink

<p>Lightweight small-sized neural networks have been developed for mechanical bearing fault diagnosis. However, the information fusion technology used in the past lightweight neural networks fail to achieve a&#xa0;balance between information utilization and the effectiveness of fusion, which has resulted in misjudgments of certain fault types. In this paper, the Two-Stage Fusion Dual-Input Residual Inception Network (TSFDIRIN) is proposed. The TSFDIRIN is primarily constructed based on the data and features information fusion strategies, aimed at effectively utilizing vibration signals from multiple sensors, further improving recognition accuracy, and reducing network volume. By combining two-dimensional stacked grayscale images with constant&#xa0;<i>Q</i> transformed time-frequency spectrograms and utilizing a&#xa0;feature fusion strategy embedded in the TSFDIRIN, rich fault information can be obtained. Besides, the inception network architecture with residual structure can enhance the stability and speed up the convergence of training process. The performance of the TSFDIRIN model is compared with that of other networks in terms of accuracy, the number of parameters, and feature extraction capability. The experimental results show that the TSFDIRIN model is able to achieve 99.7% of identification accuracy with 0.13 MB memory occupation.</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

Intelligent bearing fault diagnosis method based on lightweight network with in-depth fusion strategy

  • 期刊论文

  • 发表日期：

    2025年3月31日

查看全文
  • Xiaorui Wang,
  • Runfang Hao,
  • Kun Yang,
  • Shengbo Sang,
  • Rihui Kang,
  • Chaoqian He

摘要

Lightweight small-sized neural networks have been developed for mechanical bearing fault diagnosis. However, the information fusion technology used in the past lightweight neural networks fail to achieve a balance between information utilization and the effectiveness of fusion, which has resulted in misjudgments of certain fault types. In this paper, the Two-Stage Fusion Dual-Input Residual Inception Network (TSFDIRIN) is proposed. The TSFDIRIN is primarily constructed based on the data and features information fusion strategies, aimed at effectively utilizing vibration signals from multiple sensors, further improving recognition accuracy, and reducing network volume. By combining two-dimensional stacked grayscale images with constant Q transformed time-frequency spectrograms and utilizing a feature fusion strategy embedded in the TSFDIRIN, rich fault information can be obtained. Besides, the inception network architecture with residual structure can enhance the stability and speed up the convergence of training process. The performance of the TSFDIRIN model is compared with that of other networks in terms of accuracy, the number of parameters, and feature extraction capability. The experimental results show that the TSFDIRIN model is able to achieve 99.7% of identification accuracy with 0.13 MB memory occupation.