基于LeNet5like的迁移学习风电机组叶片覆冰故障诊断研究
Research on fault diagnosis of wind turbine icing characteristics based on LeNet5like transfer learning
吕游 1封烁 2郑茜 2邓丹 2刘吉臻2
作者信息
- 1. 华北电力大学新能源电力系统全国重点实验室 北京 102206
- 2. 华北电力大学控制与计算机工程学院 北京 102206
- 折叠
摘要
针对海上风电场和高海拔地区风机机组的叶片覆冰故障模型精度低、建模速度慢等问题,提出一种基于LeNet5like的迁移学习风电机组叶片覆冰故障诊断方法.首先,整合监控和数据采集系统的记录数据与风机覆冰情况进行预处理,建立训练数据集;其次,基于改进后的LeNet5like网络构建覆冰故障诊断模型,提取数据集中多变量间的相关性特征信息;然后,经网络参数微调迁移学习对模型进行训练,实现对其他风机覆冰故障诊断模型的快速建立;最后,经实验验证,该模型覆冰故障诊断准确率为98.90％,较无迁移模块网络训练时间缩短28 s,提升约15.91％,验证了基于LeNet5like的迁移学习风电机组叶片覆冰故障诊断方法的精确性和快速性.
Abstract
A fault diagnosis of wind turbine icing characteristics based on LeNet5like transfer learning method is proposed,to address the problems of low accuracy and slow modelling speed of icing characteristics fault models,which wind turbine units are in offshore wind farms and high altitude areas.Firstly,the recorded data from the SCADA system and the wind turbine icing situation are pre-processed to build a training dataset;secondly,the icing fault diagnosis model is constructed based on the improved LeNet5like network to extract the correlation feature information between multiple variables in the dataset;then,the model is trained by the transfer learning fine-tuning to achieve the rapid establishment of ice-cover fault diagnosis models for other wind turbines;finally,the model is experimentally validated to have an icing fault diagnosis accuracy of 98.90％,a 28 s reduction in training time and an improvement of about 15.91％over the transfer module-free network,verifying the accuracy and speed of the LeNet5like based transfer learning wind turbine blade ice-cover fault diagnosis method.
关键词
故障诊断/叶片覆冰/迁移学习/LeNet5like网络/SCADA数据Key words
fault diagnosis/blade icing/transfer learning/LeNet5like networks/SCADA data引用本文复制引用
基金项目
中央高校基本科研业务费专项(2023MS029)
出版年
2024
NETL
NSTL
万方数据