RESEARCH ON ICING DETECTION OF WIND TURBINE BLADES BASED ON Tri-SE-CNN
Aiming to address the issue of existing icing detection methods for wind turbine blades,which fail to fully utilize unlabeled data and have poor classification performance,we propose an approach called Tri-SE-CNN based on improved Tri-training and convolutional neural networks.Firstly,establish a Tri-training model based on an optimal weighted strategy to distinguish the state of unlabeled samples and expand the training set.The expanded sample set is learned by the SE-CNN model that embeds the Squeeze-and-Excite(SE)module into Convolutional Neural Networks(CNN).Combined with the strong correlation characteristics of blade icing,this paper utilizes the data from wind turbines No.15 and No.21,which were provided by the Industrial Big Data Innovation Competition Platform in China in 2017,for simulation.Additionally,data from a wind farm in Yunnan,China,are used for verification.The experimental results show that the proposed method achieves higher accuracy than CNN,support vector machine and other methods.Specifically,it achieves an accuracy of 99.96% on the No.15 wind turbine,which can provide valuable references for early warning systems regarding wind turbine blade icing.
wind turbine bladesunlabeled dataconvolutional neural networksTri-trainingsqueeze and excitation networksicing detection
NETL
NSTL
万方数据
