• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 147-156.doi: 10.3901/JME.260954

• 特邀专栏:大模型驱动的高端装备智能运维 • 上一篇    

扫码分享

基于扩散增强集成域对抗的电静液作动器零样本故障诊断方法

何显松1, 赵纯2, 李双路2, 赵孝礼1, 杨晓伟1, 宋艺博1, 温昊祥1, 姚建勇1   

  1. 1. 南京理工大学机械工程学院 南京 210094;
    2. 上海航天控制技术研究所 上海 201109
  • 收稿日期:2025-12-15 修回日期:2026-07-02 发布日期:2026-11-20
  • 作者简介:何显松,男,1999年出生,博士研究生。主要研究方向为机电液装备故障诊断与健康管理。E-mail:brownsong@njust.edu.cn;赵孝礼(通信作者),男,1991年出生,博士,副教授。主要研究方向为机电液装备故障诊断与智能运维等。E-mail:xlzhao@njust.edu.cn
  • 基金资助:
    国家重点研发计划(024YFB4709600)、国家自然科学基金(52205062,52405064)、上海航天科技创新基金(SAST2024-055)、民用航天技术预先研究(D020110)、江苏省自然科学基金(BK20220950)、中央高校基本科研业务费专项资金(30925010303)、国家市场监督管理总局重点实验室(高参数电梯智能运维)(JSTJ-IOMHL-202503)、国家科技重大专项(12124778012)、中国博士后科学基金(2025M774266)、国家资助博士后研究人员计划(GZC20252675)、中央高校基本科研业务费专项资金(30925010301)、智控实验室开放基金(ICL-2023-0305)和国防科技大学装备状态感知与敏捷保障全国重点实验室基金(6142003202415)资助项目。

Diffusion-Augmented Ensemble Domain-Adversarial Method for Zero-Shot Fault Diagnosis of Electro-Hydrostatic Actuators

HE Xiansong1, ZHAO Chun2, LI Shuanglu2, ZHAO Xiaoli1, YANG Xiaowei1, SONG Yibo1, WEN Haoxiang1, YAO Jianyong1   

  1. 1. School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094;
    2. Shanghai Aerospace Control Technology Institute, Shanghai 201109
  • Received:2025-12-15 Revised:2026-07-02 Published:2026-11-20

摘要: 针对航空、航天等领域电静液作动器(Electro-hydrostatic actuator,EHA)高可靠性的需求,导致其零样本故障诊断困难的问题,提出一种基于扩散增强集成域对抗的电静液作动器零样本故障诊断方法。在试验台与仿真模型上开展模拟故障试验,将采集到的信号在时域内进行堆叠重组,并整理为灰度图像;使用扩散模型对仿真生成的图像进行增强,构建增强仿真图像数据集;使用试验正常工况对应灰度图像与增强仿真图像数据集作为训练集,并输入到集成域对抗模型进行训练,通过调整损失函数权重来平衡训练损失,融合输出概率进一步提高诊断置信度。试验结果表明,相较于其他方法,所提方法通过物理信息与数据驱动的深度融合,在没有试验故障工况数据参与训练的情况下,实现了试验故障的精准分类,攻克了EHA故障诊断中数据稀缺与仿真-试验域偏移双重瓶颈,为复杂机电液高端装备的零样本故障诊断提供了跨域诊断的思路。

关键词: 零样本故障诊断, 扩散模型, 电静液作动器, 域对抗网络, 集成学习

Abstract: A diffusion-Augmented ensemble domain-adversarial method for zero-shot fault diagnosis of Electro-Hydrostatic Actuators (EHA) in aerospace and related high-reliability fields is proposed to address the difficulty caused by high reliability requirements. Simulated fault experiments are conducted on test benches and in simulation models, the collected signals are stacked and reorganized in the time domain, and are arranged into grayscale images. Simulation-generated images are enhanced by a diffusion model, and an enhanced simulated-image dataset is constructed. Grayscale images corresponding to normal test operating conditions and the enhanced simulated-image dataset are used as the training set and are input into an ensemble domain-adversarial model for training; training losses are balanced by adjusting loss-function weights, and output probabilities are fused to further increase diagnostic confidence. Experimental results show that, compared with other methods, precise classification of test faults is achieved without using test fault-condition data in training through deep integration of physical information and data-driven approaches; the dual bottlenecks of data scarcity and simulation-test domain shift in EHA fault diagnosis are overcome, and a cross-domain diagnostic approach for zero-shot fault diagnosis of complex electromechanical-hydraulic equipment is provided.

Key words: zero-shot fault diagnosis, diffusion model, electro-hydrostatic actuator, domain adversarial network, ensemble learning

中图分类号: