In order to improve the condition monitoring and fault diagnosis of wind turbines,a stacked noise reduction autoencoding network based on group normalization is proposed in this paper.The network is based on SCADA dat...In order to improve the condition monitoring and fault diagnosis of wind turbines,a stacked noise reduction autoencoding network based on group normalization is proposed in this paper.The network is based on SCADA data of wind turbine operation,firstly,the group normalization(GN)algorithm is added to solve the problems of stack noise reduction autoencoding network training and slow convergence speed,and the RMSProp algorithm is used to update the weight and the bias of the autoenccoder,which further optimizes the problem that the loss function swings too much during the update process.Finally,in the last layer of the network,the softmax activation function is used to classify the results,and the output of the network is transformed into a probability distribution.The selected wind turbine SCADA data was substituted into the pre-improved and improved stacked denoising autoencoding(SDA)networks for comparative training and verification.The results show that the stacked denoising autoencoding network based on group normalization is more accurate and effective for wind turbine condition monitoring and fault diagnosis,and also provides a reference for wind turbine fault identification.展开更多
基金the National Natural Science Foundation of China(51767014),2018–2021.
文摘In order to improve the condition monitoring and fault diagnosis of wind turbines,a stacked noise reduction autoencoding network based on group normalization is proposed in this paper.The network is based on SCADA data of wind turbine operation,firstly,the group normalization(GN)algorithm is added to solve the problems of stack noise reduction autoencoding network training and slow convergence speed,and the RMSProp algorithm is used to update the weight and the bias of the autoenccoder,which further optimizes the problem that the loss function swings too much during the update process.Finally,in the last layer of the network,the softmax activation function is used to classify the results,and the output of the network is transformed into a probability distribution.The selected wind turbine SCADA data was substituted into the pre-improved and improved stacked denoising autoencoding(SDA)networks for comparative training and verification.The results show that the stacked denoising autoencoding network based on group normalization is more accurate and effective for wind turbine condition monitoring and fault diagnosis,and also provides a reference for wind turbine fault identification.
