Feature Selection Method for Wind Turbine Condition Monitoring Based on Neighborhood Component Analysis
Aiming at the problem that the existing feature selection methods are difficult to select important variables from a large number of SCADA parameters,we proposed a feature variable selection method specially suitable for wind turbine condition monitoring based on the nearest neighbor component analysis.The proposed method assigns corre-sponding importance weights to each variable according to the contribution rate of each candidate variable to the regres-sion accuracy,so as to select the most important feature variables.By analyzing the characteristics of redundant varia-bles in SCADA data,we proposed a method of eliminating redundancy based on correlation coefficient matrix.The three methods of Pearson correlation coefficient,mutual information and random forest are used as comparison,and ga-ted recurrent unit network is used as model to predict the gearbox oil tank temperature.The feature selection results of the four methods are evaluated by the prediction accuracy index and residual control chart.The results show that the se-lection result of the proposed method is more intuitive,less redundant variables and higher prediction accuracy.
万方数据
维普
