现代制造工程 ›› 2026, Vol. 552 ›› Issue (9): 150-160.doi: 10.16731/j.cnki.1671-3133.2026.09.018

• 设备设计/诊断维修/再制造 • 上一篇    

基于多模态特征融合的滚动轴承故障诊断

杨小川1, 李纲1, 曹菁菁2, 吴春霖2, 李一凡2   

  1. 1 秦皇岛港股份有限公司技术中心,秦皇岛 066002;
    2 武汉理工大学交通与物流工程学院,武汉 430063
  • 收稿日期:2026-01-05 出版日期:2026-09-18 发布日期:2026-10-22
  • 通讯作者: 曹菁菁,博士,副教授,主要研究方向为机器学习和模式识别。E-mail:bettycao@whut.edu.cn
  • 作者简介:杨小川,硕士,高级工程师,主要研究方向为港口机械设备故障诊断和预测。E-mail:yangxiaochuan@portqhd.com
  • 基金资助:
    湖北省自然科学基金项目(2022CFB809)

Research on fault diagnosis of rolling bearings based on multi-modal feature fusion

YANG Xiaochuan1, LI Gang1, CAO Jingjing2, WU Chunlin2, LI Yifan2   

  1. 1 Technology Center,Qinhuangdao Port Co.,Ltd.,Qinhuangdao 066002,China;
    2 School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan 430063,China
  • Received:2026-01-05 Online:2026-09-18 Published:2026-10-22

摘要: 传统单模态滚动轴承故障诊断模型在工程实践中存在表征能力不足的问题,提出了一种基于时序和图像的多模态融合网络(Multi-Modal Fusion Network,MMFNet)。针对时序模态分支,构建时-频卷积神经网络(Time-Frequency Convolutional Neural Network,TFCNN)同步提取时域和频域特征,充分利用振动信号所包含的时序特征信息;针对图像模态分支,设计了改进的局部特征增强模块(Local Feature Enhancement,LFE)Swin-Transformer模型,即在Swin-Transformer模型前引入局部特征增强模块,以提升图像特征的局部特征表达能力;针对多模态特征融合,提出成对交叉注意力融合方法,避免特征融合时的信息冗余和特征干扰问题。为了验证模型的有效性,使用华中科技大学的滚动轴承数据集来进行实验验证,MMFNet的诊断准确率达到99.00 %以上,均高于其他单模态模型。

关键词: 多模态融合, 滚动轴承, 故障诊断

Abstract: Traditional single-modality rolling bearing fault diagnosis models suffer from insufficient representational capabilities in engineering practice. It proposes a Multi-Modality Feature Net (MMFNet) based on time-series and image data. For the temporal modality branch,a Time-Frequency Convolutional Neural Network (TFCNN) is constructed to simultaneously extract time-domain and frequency-domain features,fully leveraging the temporal information embedded in vibration signals. For the image modality branch,a Local Feature Enhancement block (LFE) Swin-Transformer is designed. This block is introduced before the Swin-Transformer to enhance the local feature representation capability of time-frequency images. For multimodal feature fusion,a Paired Bidirectional Cross-Attention (PBCA) fusion method is proposed to mitigate information redundancy and feature interference during fusion. To validate the model′s effectiveness,experiments were conducted using the rolling bearing dataset from Huazhong University of Science and Technology. MMFNet achieved diagnostic accuracies exceeding 99.00 %,outperforming other unimodal models.

Key words: multi-modal fusion, rolling bearing, fault diagnosis

中图分类号: 

版权所有 © 《现代制造工程》编辑部 
地址:北京市东城区东四块玉南街28号 邮编:100061 电话:010-67126028 电子信箱:2645173083@qq.com
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn