通信学报 ›› 2024, Vol. 45 ›› Issue (4): 185-200.doi: 10.11959/j.issn.1000-436x.2024081

• 学术通信 • 上一篇    

基于Nadam-TimeGANXGBoost的时序信号故障诊断方法

黑新宏1,2, 高苗1, 张宽1, 费蓉1,2, 邱原1,2, 姬文江1,2   

  1. 1.西安理工大学计算机科学与工程学院,陕西 西安 710048
    2.陕西省网络计算与安全技术重点实验室,陕西 西安 710048
  • 收稿日期:2023-12-07 修回日期:2024-03-12 出版日期:2024-04-30 发布日期:2024-05-27
  • 作者简介:黑新宏(1976- ),男,陕西延安人,博士,西安理工大学教授、博士生导师,主要研究方向为智能系统与安全关键系统、人工智能、大数据及其在轨道交通等系统中的应用。
    高苗 (2000- ),女,陕西宝鸡人,西安理工大学硕士生,主要研究方向为故障诊断。
    张宽 (1998- ),男,陕西西安人,西安理工大学硕士生,主要研究方向为故障诊断。
    费蓉 (1980- ),女,陕西西安人,博士,西安理工大学教授、博士生导师,主要研究方向为社区发现、随机优化算法、故障诊断。
    邱原 (1983- ),男,陕西西安人,博士,西安理工大学讲师,主要研究方向为时空数据检索、数据挖掘和自动语音识别。
    姬文江(1984- ),男,陕西延安人,博士,西安理工大学副教授、硕士生导师,主要研究方向为智能交通系统、智能和安全关键系统。
  • 基金资助:
    国家自然科学基金资助项目(62120106011);国家重点研发计划基金资助项目(2022YFB2602203)

Fault diagnosis method of timing signal based on Nadam-TimeGAN and XGBoost

Xinhong HEI1,2, Miao GAO1, Kuan ZHANG1, Rong FEI1,2, Yuan QIU1,2, Wenjiang JI1,2   

  1. 1.School of Computer Science and Engineering, Xi'an University of Technology, Xi'an 710048, China
    2.Shaanxi Key Laboratory of Network Computing and Security Technology, Xi'an 710048, China
  • Received:2023-12-07 Revised:2024-03-12 Online:2024-04-30 Published:2024-05-27
  • Supported by:
    The National Natural Science Foundation of China(62120106011);The National Key Research and Development Program of China(2022YFB2602203)

摘要:

为了提高故障诊断模型在数据不平衡场景下的诊断性能和模型泛化能力,提出了一种基于Nadam-TimeGAN和XGBoost的时序信号故障诊断方法。首先对比基于LSTM和GRU的TimeGAN模型,选取性能更优的GRU网络作为TimeGAN模型的组成单元,然后采用Nadam优化算法对TimeGAN模型的各组件进行优化,即构建Nadam-TimeGAN模型用以数据扩充,最后构建一个平衡的数据集输入XGBoost集成学习模型进行分类训练。实验选取转辙机动作电流数据集进行验证性实验,选取MFPT轴承数据集和CWRU轴承数据集进行泛化性实验,并与8种方法进行对比,结果表明,所提方法在准确率、召回率以及F1-score这3种评价指标上均高于其他方法,从而验证了所提方法在不平衡数据故障诊断方面的有效性和泛化性。

关键词: 时间序列生成对抗网络, Nesterov加速自适应矩估计, 极致梯度提升, 故障诊断, 数据增强

Abstract:

In order to improve the diagnostic performance and model generalization ability of the fault diagnosis model in data imbalance scenarios, a time series signal fault diagnosis method based on Nadam-TimeGAN and XGBoost was proposed. Firstly, the TimeGAN model based on LSTM and GRU was compared, and the GRU network with better performance was selected as the component unit of the TimeGAN model. The Nadam optimization algorithm was used to optimize the components of the TimeGAN model, that was, the Nadam-TimeGAN model was constructed for data expansion. After data expansion, a balanced data set was constructed and input into the XGBoost integrated learning model for classification training. In the experiment, the action current data set of switch machine was selected for verification experiment, the MFPT bearing data set and the CWRU bearing data set were selected for generalization experiment, and compared with eight methods. The results show that the proposed method is higher than other methods in accuracy, recall and F1-score. The experimental results validate the effectiveness and generalization of the proposed method for imbalanced data fault diagnosis.

Key words: TimeGAN, Nadam, XGBoost, fault diagnosis, data augmentation

中图分类号: 

No Suggested Reading articles found!