• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (11): 353-361.doi: 10.3901/JME.260600

• 数字化设计与制造 • 上一篇    

扫码分享

多模态机制改进差分自编码的时变速滚动轴承数据增强方法

王振亚1,2, 刘韬1,2, 吴峰崎3, 伍星2,4   

  1. 1. 昆明理工大学机电工程学院 昆明 650500;
    2. 云南省先进装备智能制造技术重点实验室 昆明 650500;
    3. 上海市特种设备监督检验技术研究院起重机械型式试验中心 上海 200062;
    4. 滇西应用技术大学 大理 671009
  • 收稿日期:2025-10-31 修回日期:2026-02-20 发布日期:2026-07-29
  • 作者简介:王振亚,男,1996年出生,博士研究生,讲师。主要研究方向数据不平衡条件下的智能装备故障诊断。E-mail:zy2022w@126.com;刘韬(通信作者),男,1980年出生,博士,教授,博士研究生导师。主要研究方向为现代信号处理理论与方法,基于人工智能理论的设备智能诊断、性能评估预测以及装备健康运维系统研发应用。E-mail:kmliutao@aliyun.com
  • 基金资助:
    国家自然科学基金(52065030)、云南省“兴滇英才支持计划”产业创新人才(KKXY202401001)、云南省教育厅科学研究基金研究生类(2024Y128)、云南省博士生服务产业科研创新培育(FWCY-BSPY2024056)资助项目。

Multimodal Mechanism Enhanced Variational Autoencoder for Variable Speed Rolling Bearing Data Augmentation Methods

WANG Zhenya1,2, LIU Tao1,2, WU Fengqi3, WU Xing2,4   

  1. 1. Electromechanical Engineering School, Kunming University of Science and Technology, Kunming 650500;
    2. Yunnan Key Laboratory of Advanced Equipment Intelligent Manufacturing Technology, Kunming 650500;
    3. Crane Type Testing Center, Shanghai Institute of Special Equipment Supervision, Inspection and Technology, Shanghai 200062;
    4. West Yunnan University of Applied Sciences, Dali 671009
  • Received:2025-10-31 Revised:2026-02-20 Published:2026-07-29

摘要: 针对差分自编码器(Variational autoencoder,VAE)在数据增强时重构质量不佳,潜在空间分布不准确的问题,提出了一种多模态机制改进VAE (Multi-modal mechanism enhanced VAE,MMEVAE)的时变速滚动轴承数据增强方法。首先,将采集到的时变速轴承振动信号进行短时傅里叶变换得到信号的时频分布矩阵,依次计算每个时间窗口上频谱的熵并进行归一化,得到归一化时频熵谱图(Normalized time-frequency entropy spectrum,NTFES)作为模型的输入;然后,引入高斯混合模型(Gaussian mixture model,GMM)重新定义潜在空间,基于贝叶斯信息准则(Bayesian information criterion,BIC)与互信息(Mutual information,MI)对潜在空间的高斯分布个数进行评估和选择,在图像重建质量和模型复杂度之间取得平衡,并基于多模态特点改进了损失函数以提高模型拟合能力;其次,在模型中嵌入自注意力机制(Self-attention mechanism,SAM),增强模型对关联信息的特征提取能力以提高图片生成质量;最后,使用真实的变转速轴承故障模拟实验验证了方法的有效性。实验结果表明,所提方法能生成高质量的虚拟数据,可有效提高“变速-数据不平衡”条件下滚动轴承的诊断精度。

关键词: 差分自编码, 多模态机制, 不平衡数据, 变转速, 故障诊断

Abstract: To address the issues of poor reconstruction quality and inaccurate latent space distribution in Variational Autoencoder (VAE) during data augmentation, a multi-modal mechanism enhanced VAE (MMEVAE) method for time-varying speed rolling bearing data augmentation is proposed. Firstly, the collected vibration signals are short-time Fourier transformed to obtain the time-frequency distribution matrix of the signals, and the entropy of the spectra in each time window is calculated and normalized in turn, so that the NTFES is obtained as the input of the model. Then, GMM is introduced to redefine the potential space, and the number of distributions of the model’s potential space models is evaluated and selected based on BIC and mutual information MI, so that the quality of image reconstruction and model complexity can be improved. Secondly, the SAM is embedded in the model to capture the correlation information between layers more effectively to enhance the quality of image generation. Finally, the effectiveness of the proposed method is verified using a real variable speed bearing fault simulation experiment. The experimental results show that the proposed method can generate high-quality virtual samples to improve the diagnostic accuracy.

Key words: variational autoencoder, multimodal mechanism, imbalanced data, variable speed, fault diagnosis

中图分类号: