发电技术 ›› 2023, Vol. 44 ›› Issue (6): 824-832.DOI: 10.12096/j.2096-4528.pgt.22178

• 发电及环境保护 • 上一篇    下一篇

基于重采样降噪与主成分分析的宽卷积深度神经网络风机故障诊断方法

刘展1, 包琰洋2, 李大字2   

  1. 1.北京能高普康测控技术有限公司,北京市 丰台区 100070
    2.北京化工大学自动化研究所,北京市 朝阳区 100029
  • 收稿日期:2023-03-29 出版日期:2023-12-31 发布日期:2023-12-28
  • 通讯作者: 李大字
  • 作者简介:刘展(1982),男,博士,高级工程师,主要从事风力发电行业基于工业物联网的核心关键设备服役能力保持与主动安全保障技术研究、产品开发及工业物联网大数据分析服务等相关工作,LiuZhan@bj-pukang.com
    包琰洋(2000),男,硕士研究生,主要研究方向为强化学习与故障诊断等,2022200727@buct.edu.cn
    李大字(1970),女,博士,教授,博士生导师,从事机器学习与人工智能、先进控制、复杂系统建模与优化等研究,本文通信作者,lidz@mail.buct.edu.cn
  • 基金资助:
    国家自然科学基金项目(62273026);工信部高技术船舶科研项目(MC-202025-S02)

Fault Diagnosis Method of Wind Turbines Based on Wide Deep Convolutional Neural Network With Resampling and Principal Component Analysis

Zhan LIU1, Yanyang BAO2, Dazi LI2   

  1. 1.Beijing Pukang Automation Technology Co. , Ltd. , Fengtai District, Beijing 100070, China
    2.Institute of Automation, Beijing University of Chemical Technology, Chaoyang District, Beijing 100029, China
  • Received:2023-03-29 Published:2023-12-31 Online:2023-12-28
  • Contact: Dazi LI
  • Supported by:
    National Natural Science Foundation of China(62273026);High-Tech Ship Research Project of Ministry of Industry and Information Technology(MC-202025-S02)

摘要:

针对数据驱动的风机故障诊断面临的数据量少、信号噪声干扰等问题,提出了一种基于宽卷积深度神经网络的故障诊断方法。该方法采用了重采样、小波阈值去噪等信号预处理方式,既增加了信息密度,又保证了信息的完整性,结合主成分分析法(principal component analysis,PCA)替代人工经验进行数据通道的选取。利用卷积神经网络的强大特征提取能力,通过较少的数据训练即可对风机机组在时域上的故障信号进行有效的特征提取,从而可以对风机进行精确的故障诊断。基于某真实风机机组数据的实验结果,验证了该方法的有效性。

关键词: 风机, 宽卷积深度卷积神经网络, 重采样, 小波阈值去噪, 主成分分析法

Abstract:

Fault diagnosis of wind turbines suffers from less training data and noises. A method based on wide deep convolutional neural network with resampling and principal component analysis was presented for the diagnosis of mechanical faults (that is the main fault component of wind turbines). The method adopted a variety of signal preprocessing methods such as resampling wavelet threshold denoising and principal component analysis to increase the information density and ensure the integrity of the information. After being trained with small amount of data, the network which has a powerful feature extraction capability could extract the fault signal in the time domain which will be further used for fault diagnosis. Experimental results were verified based on the real wind turbine data, demonstrating the effectiveness of this method.

Key words: wind turbine, wide deep convolutional neural network, resampling, wavelet threshold denoising, principal component analysis

中图分类号: