Few-shot Fault Diagnosis For Pitch System Of Wind Turbines Based On Prototypical Network With Mahalanobis Distance

As a kind of renewable energy,wind energy attracts more and more attention,and wind power has been vigorously promoted and developed.Because the wind turbine usually runs under harsh environmental conditions,the key components of the turbine are seriously damaged,especially the failure of the rotor system.Due to the increasing scale,complex structure and remote installation sites of wind turbines,the maintenance cost of wind turbines is high,which seriously affects the operation efficiency of wind power enterprises.On the other hand,due to the wide application of condition monitoring and acquisition system,the wind turbine generates massive operation data during operation,which creates good conditions for the condition monitoring and fault diagnosis of wind turbine.However,it should be pointed out that the fault history data accounts for a small proportion of the massive operation data,which brings new difficulties to the fault diagnosis of key components of wind turbines.This paper takes the variable rotor system with frequent wind turbine failures as an example.A fault diagnosis strategy based on few-shot learning is proposed,and a fault diagnosis model of the rotor system based on Mahalanobis distance prototype network is established.The specific research contents are as follows:（1）On the basis of expounding the working principle of wind power generation and explaining the basic composition and characteristics of direct drive wind generator,the working principle of variable blade system is described in detail,and the variable angle fault,variable blade torque fault and variable blade motor fault of variable blade system are analyzed;（2）Aiming to solve the problem of fault diagnosis of wind turbine rotor system,the study sets up a fault tree to analyze the source of Rotor system failure.On the basis of this,a fault diagnosis method based on Mahalanobis distance prototypical network is proposed.The prototypical is got by averaging embedding function prototype of each training sample.Besides,the study contrasts test sample and classifies the Mahalanobis distance between each type of prototype.Adam is used to optimize the network parameters.The validity of the proposed fault diagnosis model is verified by taking the western storage bearing data set of USA as an example;（3）Based on the actual operation data of a wind field in northeast China,the fault diagnosis models of variable rotor system are compared and analyzed.Principal component analysis was used to select the feature of the fault diagnosis model,and the long short-term memory neural network was used to reconstruct the variables.Several small sample learning methods,such as matching network,prototypical network and MAML,were compared with the Mahalanobis distance prototypical network proposed above to verify the accuracy of the proposed fault diagnosis method.The superiority of Mahalanobis distance is analyzed by comparing Euclidean distance,Manhattan distance,cosine distance and correlation distance.The results show that the proposed fault diagnosis model based on Mahalanobis distance prototypical network has high fault classification accuracy.