Research On Fault Diagnosis Of Wind Turbine Based On SCADA Data

Wind energy is a widely distributed,clean,low-cost renewable energy,using wind power is important for our country to achieve "carbon peak","carbon neutralization" long-term goal.In recent years,wind power industry scale continues to grow rapidly,and has been obtained in large-scale development and application.However,due to the harsh environment in which the wind turbine is located,various components of the wind turbine are prone to failure.In order to achieve fault warning early,most wind turbines are equipped with Supervisory Control And Data Acquistion(SC AD A)systems to ensure the safe and stable operation of the wind turbines.In this paper,the SCADA dataset collected during the operation of wind turbines is used to build a fault diagnosis model by introducing deep learning and transfer learning methods.The main research contents include:(1)Summarize various fault diagnosis methods for wind turbines that have been proposed and analyze their advantages and disadvantages.This paper illustrates the significant advantages of fault diagnosis method based on SCADA big dataset combined with deep learning.(2)The characteristics of SCADA dataset are analyzed,and preprocess the dataset based on the wind speed power curve.The threshold value was first set to filter the accumulated data points above the wind power curve.Then,compare the outlier detection effects of the DBSCAN algorithm,the 3Sigma criterion,and the quartile algorithm.An outlier recognition algorithm based on interval optimization is proposed.Finally,normalization and correlation analysis of characteristic variables were carried out on the dataset,and 10-dimensional variables were selected from the original 26-dimensional characteristic variables as input for the subsequent construction of the model.(3)In order to make full use of temporal and spatial correlation between characteristic variables in SCADA dataset,this paper proposes a spatio-temporal feature fusion fault diagnosis model based on attention mechanism.In the proposed model,a temporal feature extraction module and a spatial feature extraction module are designed to extract fault features,and focus loss function is used to replace the traditional cross entropy loss function to improve the unbalance effect of sample categories.During the training process,the Dropout method is introduced to suppress the overfitting phenomenon,which makes the constructed model have stronger generalization performance.During simulation experiments,datasets are separated by sliding windows as input to the model.Compared with the existing methods,this method has better performance,especially for the recognition of minority samples with higher accuracy.Finally,the t-distributed Stochastic Neighbor Embedding(t-SNE)method is used to analyze the strong feature mining ability of the model proposed in this paper,which can accurately identify between fault and normal sample features.(4)Aiming at the problem of missing fault samples and tags in the dataset collected during the actual operation of wind turbines,the idea of transfer learning was introduced.Based on the model mentioned in Chapter 3,a fault diagnosis model based on parameter transfer is proposed.In the simulation experiment,sufficient pre-training is carried out on the model in the source domain.When transfering to the target domain,four transfer strategies are used to fine-tune the model.The results show that when the source and target domains are selected appropriately,using the method of parameter migration combined with fine tuning can further improve the accuracy of fault diagnosis while reducing the time under the condition of lack of fault samples compared with no transfer learning.