Investigations On Fault Diagnosis Model Of Wind Turbine Blade Icing Based On Machine Learning

At the background of the "dual-carbon" target and the rapid development of clean energy,the installed capacity of wind turbines has been increasing yearly.Researching industrial intelligent fault diagnosis technology is of great significance in reducing the operation and maintenance costs of wind turbines.Among these faults,the icing on the blades seriously affects the operation of wind turbines,and traditional detection methods are time-consuming and expensive.In this paper,we use the actual operating data of the Data Acquisition and Monitoring System(SCADA),coupled with feature engineering and Bayesian optimization algorithms,to establish a machine learning-based fusion model to diagnose wind turbine blade icing faults.This significantly improves the accuracy and speed of icing fault diagnosis.The main research content and conclusions are as follows:(1)Regarding feature processing in the wind turbine blade icing fault diagnosis model,a high correlation mixed feature was constructed using statistical analysis methods.We proposed a polynomial feature fusion method based on a decision tree algorithm.This improves the correlation coefficient and mutual information value between the selected features and the actual icing status of wind turbine blades.Using different integrated algorithms to establish the wind turbine blade icing fault model and comparing their prediction results before and after feature engineering,we calculated the model’s F1 value(the harmonic mean of precision rate and recall rate)after feature engineering.The gradient boosting decision tree(GBDT)model increased by 2.71%,the extreme gradient boosting(XGBoost)model increased by 4.61%,and the random forest(RF)model increased by 0.65%.The results indicate that this feature fusion method can improve the accuracy of the wind turbine blade icing fault diagnosis model.(2)For hyperparameter optimization of the wind turbine blade icing fault diagnosis model,we compared the Bayesian optimization algorithm based on TPE proxy with the random grid search algorithm in the process of model tuning to reduce the computational and time cost of modeling.The results show that the Bayesian optimization based on the TPE proxy can gradually approach the optimal parameters in the search process with stronger directionality.For the blade icing fault diagnosis model with high complexity and large computational load,using the TPEbased Bayesian algorithm for hyperparameter optimization can achieve a more efficient search in a limited time,thus saving a lot of time and computing power.Moreover,with the same number of iterations in the later stage of iterative calculation,the TPE-based Bayesian optimization provided a higher best F1 value while producing better search results.It means that a more excellent combination of hyperparameters can be obtained.By using different integrated algorithms to establish the blade icing fault diagnosis model and optimizing the F1 value of single models based on TPE proxy Bayesian optimization search,we found the GBDT model increased by 13.44%,the XGBoost model increased by 7.55%,and the RF model increased by 0.23%.The results indicate that this method effectively improves the accuracy and speed of the icing fault diagnosis model.(3)Coupling feature engineering and Bayesian optimization algorithm based on TPE proxy,we constructed different blade icing fault diagnosis fusion models using voting and stacking fusion strategies.We studied the diagnostic performance of the fusion model after supplementing the original features as input to the stacking second-level learner.Our results show that in voting fusion,the F1 value of the weighted fusion model(VotingQ)was better than that of the ordinary fusion model(VotingP).In stacking fusion,using XGBoost,GBDT,and RF models as first-level learners and XGBoost as a second-level learner,and after feature enhancement,the fusion model Stacking-XGBoost achieved the best F1 value of 96.72%.Furthermore,we compared and analyzed the prediction results of the fusion model and the single model.The F1 value of the fusion model Stacking-XGBoost increased by 8%compared to the original XGBoost model,and the performance of the weighted fusion model VotingQ increased by 0.76%compared to the optimized XGBoost model.The performance of the fusion model Stacking-XGBoost increased by 1.48%.This proves that the fusion strategy further improves the generalization performance of the wind turbine blade icing fault diagnosis model.The improvement space of the stacking fusion strategy is larger.In summary,the fusion model constructed by coupling feature engineering and Bayesian optimization algorithm based on TPE proxy can significantly improve the accuracy of blade icing fault diagnosis.