arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.07054cs.AIeess.SP
收藏

双向数字孪生原型锚定与多周期学习用于少样本故障诊断

Bi-directional digital twin prototype anchoring with multi-periodicity learning for few-shot fault diagnosis

State Key Laboratory of Mechanical System and Vibration(机械系统与振动国家重点实验室);Shanghai Jiao Tong University(上海交通大学)

Pengcheng Xia, Zhichao Dong, Chengjin Qin, Qun Chao, Yixiang Huang, Chengliang Liu

AI总结:

本文提出一种双向数字孪生原型锚定与多周期学习方法,用于解决少样本故障诊断中的数据不足问题,通过元训练和测试时适应提升模型鲁棒性,实验验证其有效性。

摘要:

智能故障诊断(IFD)已作为一种确保工业机械安全性和可靠性的强大范式。然而,传统IFD方法严重依赖于大量标记数据进行训练,这在实际工业环境中往往难以获得。因此,构建物理资产的数字孪生(DT)以获取模拟数据已成为一种有前景的替代方法。然而,现有DT辅助诊断方法主要通过领域适应技术转移诊断知识,这仍然需要从目标资产中获取大量未标记数据。为了解决在仅极少量样本可用的少样本场景中的挑战,提出了一种双向DT原型锚定方法与多周期学习。具体而言,构建了一个包含在DT虚拟空间中的元训练和在物理空间中的测试时适应的框架,以实现可靠的少样本模型适应。进一步开发了一种双向双域原型锚定策略,结合协方差引导的增强以提高原型估计的鲁棒性。此外,设计了一个多周期特征学习模块,以捕捉当前信号中的内在周期性特征。基于有限元方法构建了异步电动机的DT,并在多种少样本设置和三种工作条件下进行了实验。比较和消融研究证明了所提方法在少样本故障诊断中的优越性和有效性。

英文摘要:

Intelligent fault diagnosis (IFD) typically requires abundant labeled data, which are difficult to obtain from physical assets in practical industrial applications. Digital twins (DTs) provide a promising alternative by generating labeled simulation data, while most existing DT-assisted diagnosis methods rely on domain adaptation and still require considerable unlabeled target-domain data. To enable diagnostic knowledge transfer when only a few labeled target samples are available, this paper proposes a bi-directional DT prototype anchoring framework with multi-periodicity learning for few-shot fault diagnosis. The model is first meta-trained in the DT virtual space to acquire transferable diagnostic knowledge and is then adapted to the physical asset through test-time adaptation. To improve cross-domain adaptation under extremely limited target data, a bidirectional twin-domain prototype anchoring strategy is developed to mutually constrain the source- and target-domain class representations, while covariance-guided augmentation is introduced to improve the robustness of target prototype estimation. In addition, a multi-periodicity feature learning module is designed to capture the intrinsic periodic characteristics of motor current signals. A finite-element-based DT of an asynchronous motor is constructed for experimental validation under three rotational speeds and multiple few-shot settings. The experimental results show that the proposed method outperforms representative comparison methods, while the ablation studies verify the effectiveness of the components. These results demonstrate the potential of the proposed framework for DT-to-asset diagnostic knowledge transfer under severely limited target-domain data.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑