基于无源自监督域适应的旋转机械跨域新故障诊断方法  

Cross-domain Emerging Fault Diagnosis of Rotating Machinery Using Source-free Self-supervised Domain Adaptation Network

作　　者：乐珂 李霁蒲 陈祝云 何国林[2,3] 邓书涵 李巍华 YUE Ke;LI Jipu;CHEN Zhuyun;HE Guolin;DENG Shuhan;LI Weihua(Shien-Ming Wu School of Intelligent Engineering,South China University of Technology,Guangzhou 510640;School of Mechanical&Automotive Engineering,South China University of Technology,Guangzhou 510640;Pazhou Lab,Guangzhou 510005)

机构地区：[1]华南理工大学吴贤铭智能工程学院,广州510640 [2]华南理工大学机械与汽车工程学院,广州510640 [3]琶洲实验室,广州510005

出　　处：《机械工程学报》2025年第3期67-76,共10页Journal of Mechanical Engineering

基　　金：广东省重点领域研发计划(2023B0909050007,2021B0101200004);国家自然科学基金(52275111,52075182,52205101)资助项目。

摘　　要：随着新一代人工智能技术的迅猛发展,数据驱动的智能故障诊断方法已经在机械装备中已取得广泛应用。智能诊断模型的高精度诊断通常依赖于大规模的已知标签数据。然而,机械装备在运行过程会发生不可预知的新故障,这使得已知故障数据训练的模型难以准确识别新出现的故障类型。此外,数据隐私保护限制着数据的可访问性,为诊断模型的域适应过程带来了巨大挑战。为此,提出一种基于无源自监督域适应的跨域新故障诊断方法,在源域数据仅参与模型预训练的情况下,实现对目标域中已知故障的精准分类和新故障的智能诊断。首先,利用带标签的源域样本预训练得到源域诊断模型;其次,利用基于不确定性熵的自监督伪标签技术筛选出高置信度的已知故障样本;最后,将目标域中带伪标签的已知故障样本与无标签的故障样本相结合,并通过对抗训练策略辨别新故障类型。利用实验室测试采集的汽车变速箱数据集开展跨域新故障诊断试验,对所提算法的有效性进行验证。试验结果表明,所提方法能够仅利用源域诊断模型和无标签目标域数据,在保障已知故障高精度诊断的同时,实现对目标域中新故障的精准诊断。With the rapid advancement of next-generation artificial intelligence technologies,data-driven intelligent fault diagnosis methods have found widespread applications in mechanical equipment.High-precision diagnosis of intelligent diagnostic models typically relies on a large volume of labeled data.However,unpredictable new faults may occur during the operation of mechanical equipment,which makes it difficult to adopt the model trained on known samples to accurately identify newly occurring faults.Furthermore,data privacy restricts the accessibility of data,adding substantial complexity to the domain adaptation process for diagnostic models.To address these challenges,a source-free self-supervised domain adaptation network is proposed for cross-domain emerging fault diagnosis of rotating machinery,which enables diagnosis emerging fault in target domain without access to source domain.First,a source domain fault diagnosis model is established using labeled source samples.Subsequently,a self-supervised pseudo-labeling technique based on uncertainty information entropy is utilized to acquire a target domain dataset with high-quality pseudo-labels.Finally,we merge the pseudo-labeled fault dataset with unlabeled fault dataset,then train the model through an adversarial training strategy to realize emerging fault detection.The effectiveness of the proposed method is validated through experiments conducted on an automotive gearbox dataset.Experimental results show that the proposed method can accurately detect emerging faults while maintaining the performance in classifying known fault types.

关 键 词：智能故障诊断 无源域适应 新故障 自监督学习 迁移学习 

分 类 号：TG156[金属学及工艺—热处理]