Wind Turbine Fault Diagnosis Based On SCADA Data And Deep Learning

With the continuous reduction of fossil energy and the approaching "double carbon" target,clean energy has been developed rapidly.Among them,the use of wind energy resources has also grown,and the installed capacity of wind turbines has been rising year by year.However,the working environment of wind turbines is often very harsh,which can easily lead to potential failures.Once a fault is not handled in time,it will affect the normal operation of wind turbines and even cause serious accidents such as downtime and equipment damage.In order to avoid downtime and prevent equipment damage,effective fault diagnosis of wind turbines is required.The wind turbines are equipped with many sensors,and the SCADA system can use these sensors to collect a large amount of wind turbine operation data,including the fault data of the wind turbine.There is a large amount of fault information hidden in these data,and how to effectively use these fault information for fault diagnosis is of great practical significance to ensure the normal operation of wind turbines and reduce the operation and maintenance costs of wind farms.Based on wind turbine SCADA data,this thesis uses principal component analysis(PCA)to reduce the dimensionality of data and eliminate linear correlation between data,and inputs the dimensionality-reduced data into a one-dimensional convolutional neural network(1DCNN)to fully learn the fault features,and then outputs the diagnosis results in response to the problems of large data dimensionality and high redundancy.By comparing with other dimensionality reduction algorithms based on the same data samples,the results prove that PCA has better dimensionality reduction effect.Compared with the conventional fault diagnosis model,the PCA-1DCNN model has better fault diagnosis resultsIn order to improve the accuracy of fault diagnosis and solve the problems of low data volume and incomplete and inaccurate feature extraction,this thesis uses the sliding window method to reuse the existing fault data in order to expand the fault samples.Then,the input samples are zero-mean normalized,and the weights of different samples and different parameters are adaptively adjusted using a compressed excitation network(SEnet).After that,to avoid the loss of early features of the network,the 1DCNN network structure is improved by introducing a residual network,an upsampling layer,a multiscale module and a hedging module to obtain multiscale features and differentiation features of the faults and to increase the network width.In addition,a batch normalization layer is introduced to regulate the data feature distribution and reduce the number of iterations to reach the optimum.The results of the diagnostic examples show that the constructed compressed incentive network weighted multi-scale hedged one-dimensional convolutional neural network(SE-MSH-1DCNN)is more accurate than the PCA-1DCNN model,and the SE-MSH-1DCNN model has better fault diagnosis capability compared with other advanced fault diagnosis methods.By testing the generalization capability of the model,the results show that the fault diagnosis accuracy of SE-MSH-1DCNN model is above 99% under different random states and five-fold cross-validation.The comparison results on different data sets confirm that the SE-MSH-1DCNN model outperforms all the other comparison models.To address the problem that the target domain does not have enough data to train a highly available diagnostic model,this thesis proposes to adopt the theory and method of migration learning to solve this practical problem.The specific approach is as follows:similar WTG fault diagnosis related knowledge is migrated to the target wind turbine to improve the fault diagnosis accuracy of the target WTG.The SE-MSH-1DCNN is used as the base model,and similar WTG data is used for training,then the parameters of the trained model are fine-tuned using the target domain data,and finally the finetuned parameters are applied to the target domain data to improve the fault diagnosis accuracy of the target wind turbine.By comparing the SE-MSH-1DCNN model with other models,the SE-MSH-1DCNN model with migration learning outperforms the comparison model in terms of accuracy and generalization ability,thus demonstrating that migration learning can be applied to this research area and can be an effective method to solve such problems.