| At present,the wind power generation technology is developing rapidly all over the world.However,the traditional maintenance mode is difficult to meet the demand when the wind turbine gearbox break down,due to the design defects,remote location and poor working environment of wind turbines.Therefore,the maintenance of wind turbines requires to be transformed to big data and intelligence.Deep learning methods can automatically extract the abstract features from massive vibration signals and is able to be a reliable analysis tool for intelligent maintenance scenarios.As wind turbine gearboxes are interfered by complex internal structures,long transmission chains and bad operating environments,vibration signals have the characteristic of information coupling,which limits the recognition rate of deep learning fault diagnosis.At the same time,the inconsistent probabilities of various faults for wind turbine gearboxes in long-term service lead to unbalanced samples among different fault categories,which make deep learning models easy to over fit and performance degradation.In view of the above problems,this paper studies fault diagnosis methods of wind turbine gearboxes based on the one-dimensional fully decoupled network according to the vibration signals characteristics of information coupling and sample imbalance.The main research of the paper is listed as follows:(1)Aiming at the problem of fault feature information caused by the long transmission chain,many moving parts and bad working conditions for wind turbine gearboxes,the one-dimensional fully decoupled network is proposed for fault diagnosis.In this method,the hyperball decoupled operator and the hypersphere decoupled operator are used to replace the inner product of convolution layers and fully connected layers in a convolution neural network,which are used to enhance the response of fault features.At the same time,the output of the last fully connected layer is amplified to punish the misclassification of the model,so that the recognition rate of fault diagnosis is improved for wind turbine gearboxes.(2)Aiming at the problem of imbalanced data for the actual operation of wind turbines,the one-dimensional fully decoupled network with sampling-risk-class-balanced loss is proposed for fault diagnosis.This method further improves the loss function on the basis of one-dimensional fully decoupled network.At first,the number of effective samples according is estimated to the data sampling model based on the maximum entropy principle.And then,the correction coefficient is constructed from the variation risk.Finally,the balanced loss weight is derived.Therefore,the problems of over fitting and model performance are solved degradation for the deep learning method under imbalanced data.(3)Under the current trend of the big data governance in the wind power industry,a system prototype is designed for intelligent wind turbine gearbox fault diagnosis,in which the application of fault diagnosis and visualization are implemented.At the end of this paper,the research contents and the practical application are summarized and the limitations are prospected. |
