A Data-Driven Method For Detecting Ice On Wind Turbine Blades

Ice cover on wind turbine blades can alter wind turbine power generation,torque and blade life,affecting wind farm power generation planning and economic efficiency.Traditional wind turbine blade ice detection is achieved by adding external sensors and manual inspection,which are limited by domain knowledge,personnel expertise and cost,and are prone to misjudgment.With the application of sensor technology in wind turbines and the continuous development of data-driven methods,it has become possible to detect paddle ice by learning potential information from historical data through artificial intelligence.In this paper,the wind turbine blade ice overlay detection is abstracted as a multivariate time series classification problem,and the SCADA data from a real operation of a wind farm in Inner Mongolia is used as the research object to study the wind turbine blade ice overlay detection problem.The main research elements are as follows:(1)A supervisory model for wind turbine blade icing detection based on Spearman’s correlation coefficient-Attention-LSTM is proposed with fully labeled wind turbine icing data as the research object.In this paper,we use Spearman correlation coefficient to correlate each parameter of wind turbine,remove the features with high consistency,so as to reduce the feature dimension,avoid the repeated extraction of features by the model and reduce the redundant information,and finally use Attention-LSTM to model the multivariate sequence of remaining features.The experimental results show that compared with other data-driven models,the model in this paper has better results in accuracy,recall,F1,and MCC,and the model is 99%accurate for judging the icing status of wind turbine blades.(2)An unsupervised wind turbine blade icing detection model TCAD is proposed with unlabeled wind turbine icing data.the model uses Transformer and Resnet to learn the global and local features of the sequences,and is constrained with reconstruction loss.Transformer and Resnet are used to learn the global and local features of the sequences and constrain the learning of rich global local representations using the difference in reconstruction and the difference in global local representations.In addition,an anomaly score based on the difference between global and local representations is proposed in this paper.The experimental results show that the model can effectively handle the wind turbine blade icing fault detection problem and has advanced performance on publicly available imbalanced data sets with some generalization capability.(3)Establishing the web-side deployment of the model.In this paper,web-side deployment of the deep learning model is implemented using python,CSS,HTML and Javas Cript.CSS,HTML and Javas Cript are used to implement the front-end page development.My SQL is used as the database to manage the account information and data.Convert the trained model from Pytorch framework to web-usable format.Deployed with web as a window.Implemented the result visualization reality.3.build the local web-side deployment of the model.In this paper,web-side deployment of the deep learning model is implemented using python,CSS,and HTML.CSS and HTML are used to implement front-end page development.My SQL is used as the database to manage the account information and data.Convert the trained model from Pytorch framework to web-usable format.Local deployment with web side as a window.Implemented training process and results visualization reality.