Weak fault early warning method for wind turbines under random operating conditions based on f-AnoGAN
WANG Weiyu
WEI Jiada
ZHANG Hai
WANG Sijia
ZHANG Pei
HE Jianjun
Abstract：[Objective]To address the issues that features of weak fault vibration signals in wind turbine gearboxes under random operating conditions are difficult to extract,and that traditional methods suffer from coarse operating condition division and limited fault recognition effectiveness,a fault early-warning model based on fast unsupervised anomaly detection with generative adversarial network(f-AnoGAN)was constructed.[Methods]Firstly,a 36-layer autoencoder was designed as the model generator,where the encoder extracted global features and the decoder completed signal reconstruction to learn the global distribution of normal signals.Secondly,a convolutional neural network was built as the discriminator to extract local features and determine the authenticity of samples,thereby enhancing the capability to capture local details of weak faults.Thirdly,by integrating the mean squared error reconstruction loss of the autoencoder and the binary cross-entropy adversarial loss of the generative adversarial network,a weighted total loss function was constructed to achieve the collaborative optimization of global and local features.Finally,using the MCC5-THU gearbox multi-mode fault dataset,model training and testing were completed under the parameters of a learning rate of 0.000 2,100 epochs,and a batch size of 32.[Results]The results demonstrate that the area under the receiver operating characteristic curve of the model reaches 0.93,and the fault discrimination accuracy is 96.18%.The latent space can effectively separate normal and abnormal data,and the loss tends to stabilize in the later stages of training.Without needing pre-set operating condition division,this model can directly learn normal patterns from time-frequency diagrams to achieve accurate early warning of weak faults under random operating conditions,which compensates for the shortcomings of traditional time-frequency domain analysis and provides a reference for online early warning of weak faults in wind turbine gearboxes.
Keywords：f-AnoGANWind turbineWeak fault early warningVibration signalRandom operating condition
Publication Date：2026-07-15
Online Publishing Date：2026-09-12(First online date of this platform, not the publication date of the document)
Pages：9( 154-162 )