Page 279 - 《振动工程学报》2026年第2期

第 39 卷第 2 期 振 动 工 程 学 报 Vol. 39 No. 2
2026 年 2 月 Journal of Vibration Engineering Feb. 2026

基 于 多 尺 度 残 差 动 态 域 适 应 网 络 的

不 同 工 况 下 转 子 故 障 诊 断 方 法

向 玲 ， 王 宁 ， 邴汉昆 ， 胡爱军 ， 韩忠泉 1
1,2
1,2
1
1
（1. 华北电力大学机械工程系，河北 保定 071003； 2. 河北省电力机械装备健康维护与失效预防重点实验室，河北 保定 071003）
摘要：不同工况下转子数据分布差异大，导致传统故障诊断模型精度低。本文提出了一种基于多尺度残差动态域适应网络
（multi-scale residual dynamic domain adaptation network，MsRDDA）的不同工况下转子故障诊断方法，用于解决源域样本有标签
而目标域样本无标签的问题，实现不同工况间的无监督迁移诊断。将采集得到的一维时域信号进行分割，并通过短时傅里叶
变换 (short-time Fourier transform，STFT) 将其转换成具有时频特征的二维图像；提出一个融合多尺度卷积和可分离卷积的多尺
度残差网络，该网络由多尺度卷积层作为输入层提取浅层特征，通过 4 个改进残差模块提取深层特征，保证提取故障特征多样
性的同时避免网络因深度的增加而产生梯度消失的问题；将动态分布域适应策略引入多尺度残差网络中，根据平衡因子动态
衡量边缘分布和条件分布的重要性，对齐特征分布，提高模型的迁移诊断性能。运用所提方法对转子试验台采集得到的数据
进行跨工况迁移诊断试验，并与其他传统迁移模型进行对比，验证了该方法的有效性和优越性。
关键词： 故障诊断；转子；迁移学习；残差网络；动态域适应
中图分类号：TH133;TH165.5 文献标志码：A DOI：10.16385/j.cnki.issn.1004-4523.202312029

Rotor fault diagnosis method based on multi-scale residual dynamic domain

adaptive network under different working conditions

1
1,2
1
1,2
XIANG Ling ，WANG Ning ，BING Hankun ，HU Aijun ，HAN Zhongquan 1
（1.Department of Mechanical Engineering，North China Electric Power University，Baoding 071003，China；
2.Hebei Key Laboratory of Electric Machinery Health Maintenance & Failure Prevention，Baoding 071003，China）
Abstract：The distribution of rotor data collected under different working conditions is very different，which leads to the low accuracy of
traditional fault diagnosis model. This paper proposes a rotor fault diagnosis method based on multi-scale residual dynamic domain adaptation
network (MsRDDA) under different working conditions. It is used to solve the problem that the source domain samples have labels but the
target domain samples have no labels， and realize unsupervised migration diagnosis between different working conditions. The one-
dimensional time domain signal collected by the rotor test bench is segmented by short-time Fourier transform (STFT) and converted into a two-
dimensional image with time-frequency domain characteristics. A multi-scale residual network combining multi-scale convolution and separable
convolution is proposed. The multi-scale convolution layer is used as the input layer to extract shallow features，and four improved residual
modules are used to extract deep features to expand the network sensitivity field，so as to ensure the diversity of fault features and avoid the
gradient disappearance caused by the increase of network depth. The dynamic distribution domain adaptation strategy is introduced into the
multi-scale residual network，the importance of edge distribution and conditional distribution is dynamically measured according to the balance
factor，and the feature distribution between the source domain and the target domain is aligned to improve the performance of the migration
diagnosis model.The proposed method was applied to conduct cross-condition transfer diagnosis experiments on the data collected from the
rotor test bench, and was compared with other traditional transfer models. The effectiveness and superiority of this method were verified.
Keywords：fault diagnosis；rotor；transfer learning；residual networks；dynamic domain adaptation

转子在旋转机械设备运行过程中起着十分关键 伤亡。因此，及时对转子的运行状态进行直接高效
的作用，转子发生故障会影响到整个装备的正常运 的故障诊断，并采取相应维修措施，对保障机械安
行，造成巨大经济损失的同时还有可能带来人员的 全、可靠地运行具有十分重要的意义。

收稿日期：2023-12-13；修订日期：2024-02-04
基金项目：国家自然科学基金资助项目（52075170，52175092)；河北省自然科学基金资助项目 (CXZZBS2022154)