现代制造工程 ›› 2026, Vol. 552 ›› Issue (9): 136-143.doi: 10.16731/j.cnki.1671-3133.2026.09.016

• 设备设计/诊断维修/再制造 • 上一篇    下一篇

基于跨设备迁移学习的转盘轴承寿命预测

赵文达, 潘裕斌, 洪荣晶, 陈捷   

  1. 南京工业大学机械与动力工程学院,南京 211816
  • 收稿日期:2025-07-21 出版日期:2026-09-18 发布日期:2026-10-22
  • 通讯作者: 潘裕斌,博士,副教授,主要研究方向为高端装备智能运维。E-mail:panyb@njtech.edu.cn
  • 作者简介:赵文达,硕士研究生,主要研究方向为机械设备状态监测。洪荣晶,博士,教授,主要研究方向为智能装备制造技术。陈捷,博士,教授,主要研究方向为设备故障诊断与健康监测。
  • 基金资助:
    国家自然科学基金项目(52205106)

Life prediction of slewing bearing based on cross-device transfer learning

ZHAO Wenda, PAN Yubin, HONG Rongjing, CHEN Jie   

  1. School of Mechanical and Power Engineering,Nanjing Tech University,Nanjing 211816,China
  • Received:2025-07-21 Online:2026-09-18 Published:2026-10-22

摘要: 转盘轴承作为大型机械设备中的核心传动部件,其寿命预测对保障工业装备安全生产、优化维护策略具有重要价值。针对转盘轴承数据稀缺且跨工况分布差异导致的寿命预测精度不足问题,提出一种跨设备迁移学习的转盘轴承寿命预测方法。首先通过自适应噪声的完备鲁棒局部均值分解(Complete Ensemble Robust Local Mean Decomposition with Adaptive Noise,CERLMDAN)方法实现振动信号降噪与多维度特征提取,继而采用网格搜寻平衡分布自适应(Grid Search-based Balanced Distribution Adaptation,GSBDA)算法,利用最大均值差异(Maximum Mean Discrepancy,MMD)构建双循环优化机制,动态平衡源域-目标域边缘分布与条件分布,最终通过编码器-解码器架构的Seq2Seq-BiGRU-Attention网络捕捉时序退化特征。实验表明,该方法在单源域迁移场景下较传统迁移方法的平均绝对误差(Mean Absolute Error,MAE)和均方误差(Mean Squared Error,MSE)降低了50 %~80 %;在多源域融合策略中使不同工况轴承数据叠加后较传统迁移方法的均方误差和平均绝对误差均降低了20 %~60 %。

关键词: 转盘轴承, 寿命预测, 迁移学习, 数据分布自适应, Seq2Seq神经网络架构

Abstract: As a core transmission component in large-scale mechanical equipment,the life prediction of slewing bearings is of significant value for ensuring the safe production of industrial equipment and optimizing maintenance strategies. Aiming at the problem of insufficient life prediction accuracy caused by the scarcity of slewing bearing data and the distribution differences across various operating conditions, it proposes a cross-device transfer learning-based approach for slewing bearing life prediction. Firstly,the Complete Ensemble Robust Local Mean Decomposition with Adaptive Noise (CERLMDAN) method is employed to achieve vibration signal denoising and multi-dimensional feature extraction. Subsequently,the Grid Search-based Balanced Distribution Adaptation (GSBDA) algorithm is adopted,utilizing the Maximum Mean Discrepancy (MMD) to construct a dual-loop optimization mechanism that dynamically balances the alignment of the marginal and conditional distributions between the source domain and the target domain. Finally,the Seq2Seq-BiGRU-Attention network with an encoder-decoder architecture is utilized to capture temporal degradation features. Experiments demonstrate that this method reduces the Mean Absolute Error (MAE) and Mean Squared Error (MSE) by 50 % to 80 % compared to traditional transfer learning methods in single-source domain transfer scenarios. In multi-source domain fusion strategies,when data from bearings under different operating conditions are superimposed,it reduces the MSE and MAE by 20 % to 60 % compared to traditional transfer learning methods.

Key words: slewing bearing, life prediction, transfer learning, data distribution adaptation, Seq2Seq neural network architecture

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