Research On Condition Monitoring Of Wind Turbine Drive Train Based On Autoencoder Model

Due to the advantages of mature technology and low power generation cost,wind power generation has attracted more and more attention in recent years and is in the stage of rapid development.Drive chain is an important transmission device to realize the conversion of wind energy to electric energy in the working process of wind turbine.The drive chain runs for a long time under the bad working conditions of variable speed and variable load,which is prone to various types of faults.Therefore,it is necessary to study the condition monitoring technology of the drive chain of wind turbines,so as to find out the abnormlities and early failures of the drive chain in time,and reduce the number of maintenance and maintenance costs of wind turbines.To solve the problems of large amount of data,strong noise and complex coupling relationship amng variables of wind turbines,this paper adopts the stacked denoising autoencoder network model in deep learning to solve the above problems.Based on the two characteristics of the deep network,the global and local state monitoring of the drive chain of wind turbines is realized.The specific research contents are as follows:(1)In overall state monitoring model of wind turbine drive chain,due to the shallow autoencoder network model of wind turbine drive chain modeling such a complex problem expression ability is limited,so set up the stacked denoising autoencoder network model can get more wind turbines with strong coupling between data contain nonlinear expression.In order to prevent false alarm caused by noise data and caused by direct setting of reconstruction error threshold,the exponential weighted moving average control chart is used to set the monitoring threshold to accurately judge whether the drive of wind turbines has a fault and when the fault occurs.(2)Aiming at the problem that the overall state monitoring model can find the abnormal drive chain but cannot determine the abnormal sub-parts of the drive chain,the local state monitoring model of wind turbine drive chain is further proposed to determine the abnormal sub-parts and maintain them accordingly to reduce the consunption of manpower and material resources.In the local state monitoring model of wind turbine drive chain,the temperature of drive chain conponents is predicted by using long and short time memory network.Considering that the stacked denoising autoencoder network has the function of data conpression and dimensionality reduction,and can remove the redundancy and noise in the multivariable input data of wind turbine,the combination of the stack denoising autoencoder network with short and long time memory network can improve the prediction accuracy of short and long time memory network.Through the sequential probability ratio test method to analyzethe temperature prediction residual error of the drive chain,the abnormal trend of the temperature of the drive chain parts can be found in time.