| 摘 要: 大量机组的老化问题导致原有故障预警系统失效,为实现老旧机组的技改增效,提出一种基于迁移学习的老化机组故障预警方法。首先,分析老化问题对机组数据分布的影响,并利用MIC(Maximal Information Coefficient)最大互信息系数和长短期记忆神经网络(Long Short-Term Memory,LSTM))构建多年度机组故障预警模型,探究老化问题对现有故障预警技术的影响;其次,基于模型的迁移学习提出老化机组故障预警方法,在保留原模型大量信息的基础上,高效且快速地解决老化问题导致的原模型失效问题。实际风场数据采集及监控系统(Supervisory Control and Data Acquisition,SCADA)的数据验证表明,所提老化机组故障预警方法与重新训练的模型均能在故障发生前46 h发出预警,此外,所提方法大幅缩减了模型的构建时间和资源。 |
| 关键词: 风电机组;老化机组;迁移学习;故障预警 |
|
中图分类号: TP206.3
文献标识码: A
|
| 基金项目: 国家自然科学基金(51975535)?,国家重点研发计划(2021YFB3301601) |
|
| A Fault Warning Method for Aging Wind Turbine Based on Transfer Learning |
|
WANG Xutao, CHEN Huanguo, TAO Hanyu, YANG Lei, GAO Xiangchong
|
(Zhejiang Sci-tech University, Zhejiang Province's Key Laboratory of Reliability Technology f or Mechanical and Electronic Product, Hangzhou 310018, China)
202130605302@mails.zstu.edu.cn; hgchen@zstu.edu.cn; taohanyu@mails.zstu.edu.cn; 202230503288@mails.zstu.edu.cn; 202230503142@mails.zstu.edu.cn
|
| Abstract: The aging problem of a large number of wind turbines has led to the failure of the original fault warning system. In order to achieve technological transformation and efficiency improvement of the aging wind turbines, this paper proposes a fault warning method based on transfer learning for aging wind turbines. Firstly, the impact of aging on the data distribution of wind turbines is analyzed, and the Maximum Information Coefficient (MIC) and Long Short-Term Memory (LSTM) neural network are used to construct a multi-year wind turbines fault warning model, exploring the impact of aging on existing fault warning technologies. Secondly, the aging wind turbine fault warning method is proposed based on model transfer learning, which efficiently and quickly solves the problem of original model failure caused by aging while retaining a large amount of information from the original model. The data validation of the Supervisory Control and Data Acquisition (SCADA) system for actual wind farms shows that both the proposed aging wind turbine fault warning method and the retrained model can issue warnings 46 hours before the fault occurs. In addition, the proposed method significantly reduces the construction time and resources of the model. |
| Keywords: wind turbines; aging wind turbines; transfer learning; fault warning |
