Intelligent Fault Diagnosis Of Wind Turbine Drivetrain Based On Deep Learning

Wind energy is one of the fastest developing renewable energy sources in recent years.Condition monitoring and fault diagnosis can effectively improve the operation reliability of wind turbines.Intelligent fault diagnosis based on deep learning reduces the operation and maintenance cost of wind turbines and can better adapt to the demand of big data.The early fault diagnosis of wind turbines can avoid heavy losses when they operate in variable conditions and harsh environments for a long time.For fault diagnosis of new wind farms,few fault samples have been obtained,or even some fault samples are missing,which leads to serious sample imbalance.The new wind farms are different from the existing wind farms in structure.The fault diagnosis model trained on the data of the existing wind farms has poor applicability and low diagnosis accuracy for the new wind farms.Aiming at the above problems,the main research work of this paper is summarized as follows:Aiming at the problem of early fault warning and location of wind turbines based on SCADA data,this paper combined sparse dictionary learning to improve the adversarial variational autoencoder.The model combines the advantages of high quality generation of generative adversarial networks and the advantages of high quality posterior distribution learning of variational autoencoders,and at the same time uses dictionary learning to extract the intrinsic features of original data.Taking three groups of faulty wind turbines as cases,and comparing with some common algorithms,the accuracy and stability of the proposed model in anomaly detection are verified.At the same time,a fault location method based on the residual errors of SCADA parameters is proposed.Aiming at the problem of few fault diagnosis samples in newly built wind farms,two few-shot fault diagnosis models for wind turbines were proposed by improving model-independent meta-learning and momentum contrast learning.By adding basic model pre-training and fine-tuning,the accuracy of few-shot diagnosis is improved.Through the fault cases of generator bearing and gearbox,the classification accuracy of the proposed model is compared with the existing model in the aspects of 5-shot and 1-shot.The accuracy is improved by about 10%compared with the existing model.Aiming at the problem of missing fault samples in newly built wind farms,three ideas are adopted:(1)Improve the conditional generation adversarial network and capsule network to realize generalized zero-shot fault diagnosis of wind turbines.Fault feature frequency is used to construct semantic attribute features for zero-shot learning.Based on the semantic attributes of each fault,the missing fault is generated using health data,and the proposed model is fine-tuned by establishing pseudo labels through the test set to solve the class-level overfitting problem.Taking the bearing data of western reserve university and the bearing data of wind turbine generator as examples,the unknown fault category is diagnosed.(2)Generalized zero-shot fault diagnosis of wind turbines is realized based on spatial embedding,and semantic attribute features are established based on fault feature frequency.By mapping samples and semantics to the same space,and alline them with label space,discriminator and contrast loss are introduced to improve classification accuracy and alleviate class-level overfitting problem.In the training phase,the mapping is learned from the samples and prototypes of visible classes to establish the relationship between visual space and semantic space.In the test phase,the semantic prototype of the invisible class is used to match the samples through the learned model to realize the diagnosis of the unknown class.Taking the data of wind turbine generator bearing as a case,the diagnosis results with one unknown class and two unknown classes are analyzed.(3)Improve model-agnostic meta-learning to realize fault transfer of wind turbines.The model includes two model-agnostic meta-learners and two classifiers,which diagnose the source domain and target domain respectively.The cross-entropy loss updates the source domain model,and the maximum mean difference between the source domain and the target domain updates the target domain model.The gearbox data of wind turbine is used as a case to realize the fault diagnosis of wind turbine.At the end of this paper,the main conclusions and the prospect of further research are given.