中国机械工程 ›› 2025, Vol. 36 ›› Issue (9): 1905-1915.DOI: 10.3969/j.issn.1004-132X.2025.09.001

• 机械基础工程 •    

基于多维复向特征融合与CNN-GRU的转子不平衡量识别方法

王坚坚1,2(), 廖与禾1,2(), 杨磊1,2, 薛久涛1,2   

  1. 1.西安交通大学现代设计及转子轴承系统教育部重点实验室, 西安, 710049
    2.西安交通大学陕西省机械产品质量保障与诊断重点实验室, 西安, 710049
  • 收稿日期:2024-06-19 出版日期:2025-09-25 发布日期:2025-10-15
  • 通讯作者: 廖与禾
  • 作者简介:王坚坚,男,1999年生,硕士研究生。研究方向为转子数字孪生、转子不平衡量无试重识别、转子-轴承系统智能诊断。E-mail:wangjianjian@stu.xjtu.edu.cn
    廖与禾*(通信作者),男,1973年生,副教授、博士研究生导师。研究方向为转子-轴承系统故障诊断、振动信号处理、早期故障诊断、智能诊断、转子动平衡。E-mail:yhliao@mail.xjtu.edu.cn
  • 基金资助:
    国家重点研发计划(2019YFB1311903);国家自然科学基金(51575424)

Rotor Unbalance Recognition Based on Multidimensional Complex Feature Fusion and CNN-GRU

Jianjian WANG1,2(), Yuhe LIAO1,2(), Lei YANG1,2, Jiutao XUE1,2   

  1. 1.Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System,Xi’an Jiaotong University,Xi’an,710049
    2.Shaanxi Key Laboratory of Mechanical Product Quality Assurance and Diagnostics,Xi’an Jiaotong University,Xi’an,710049
  • Received:2024-06-19 Online:2025-09-25 Published:2025-10-15
  • Contact: Yuhe LIAO

摘要:

现有的无试重不平衡量识别算法采用优化算法框架,通过大量迭代运算以逐步逼近最优解,这类策略普遍收敛速度迟缓且易陷入局部极值。为此,利用神经网络直接学习并解析不平衡振动响应与不平衡量之间的复杂映射关系,进而实现不平衡量的高精度识别。通过转子动力学模型进行仿真,构建了带标签的足量不平衡振动数据集。针对不平衡数据的多维复向特性,设计了一种特征融合机制。核心算法层面,结合卷积神经网络(CNN)与门控循环单元(GRU)构建了CNN-GRU混合模型,其中,CNN部分负责从振动数据中提取局部空间特征,GRU负责捕捉振动数据中的时序依赖关系,通过整合空间与时间域的信息,显著增强了模型的泛化能力和识别精度。测试集数据和实验台实验的不平衡量识别结果表明,所提方法可以准确预估识别转子的不平衡量,为无试重现场动平衡提供迅速准确的指导。

关键词: 转子, 无试重, 不平衡量识别, 卷积神经网络-门控循环单元, 多维复向特征融合

Abstract:

The existing unbalance identification algorithm without trial weight adopted an optimization algorithm framework and approximated the optimal solution through numerous iterative operations. However, such strategies typically faced the limitations of slow convergence speed and the tendency to fall into local extrema. Therefore, neural networks were used to directly learn and analyze the complex mapping relationship between unbalance vibration response and unbalance, thus realizing high-precision unbalance identification. A sufficient unbalance vibration dataset with labels was constructed by simulating the rotor dynamics model. A feature fusion mechanism was designed to address the multi-dimensional complex-valued characteristics of unbalanced data. At the core algorithm level, a CNN-GRU hybrid model was constructed. In this model, CNN was responsible for extracting local spatial features from vibration data, while GRU captured temporal dependencies within the vibration data. By integrating information from both spatial and temporal domains, the model’s generalization ability and recognition accuracy were significantly enhanced. The unbalance recognition results of test set data and experimental bench demonstrate that this method may accurately predict the unbalance of the rotors, providing a rapid and accurate guide for dynamic balancing in the field without trial weights.

Key words: rotor, without trial weight, unbalance identification, convolutional neural network-gated recurrent unit (CNN-GRU), multidimensional complex feature fusion

中图分类号: