Fault Diagnosis Method Of Wind Turbine Gearbox Based On Few-shot Constraint Deep Learning

With the rapid development of the global wind power industry,wind turbines installed in the period of high-speed growth have gradually stepped out of the warranty period,and the fault diagnosis of wind turbines has been paid more and more attention.Among them,the downtime caused by wind turbine gearbox failure is the longest,and the loss is enormous.Therefore,the fault diagnosis of wind turbine gearbox has excellent theoretical and practical significance.However,the fault vibration signal of wind turbine gearbox usually has the characteristics of non-stationary.Traditional signal processing and machine learning are often not competent,and its human participation is high,so it is challenging to realize intelligent fault diagnosis.Deep learning has been paid more and more attention because of its strong ability of non-linear feature extraction.However,the traditional deep learning-based fault diagnosis accuracy relies heavily on a large number of labeled fault data.However,some wind farms lack fault sample data due to the lack of information management in early wind farms in practical application.The fault correspondence analysis report especially leads to the lack of fault sample number and label number of wind turbine gearbox.When the number of fault samples and tags is small,the deep learning model’s generalization ability is poor,and it is easy to produce overfitting.Given the above problems,this paper studies the fault diagnosis method of wind turbine gearbox based on small sample constrained deep learning from two aspects: the lack of samples and the lack of labels.The main contents of this paper are as follows:This paper proposes a mixed self-attentive prototype network for wind gearbox fault diagnosis to address the lack of wind gearbox samples in some wind farms.First,this method maps the vibration signal to the metric space of fault features by the prototype network.Second,the mixed self-attentive module is constructed by matrix fusion using the position self-attentive mechanism and the channel self-attentive mechanism to establish the global dependence of the original vibration signal,obtain more discriminative feature information,and learn the metric prototypes for each health state of the wind gearbox.Finally,the metric classifier performs the pattern recognition to achieve the wind gearbox’s fault diagnosis with the scarce number of samples.Finally,pattern recognition is performed by a metric classifier to achieve fault diagnosis of wind turbine gearboxes under the lack of sample quantity.A large number of unlabeled data is cheap and easier to obtain to solve the lack of labels for wind turbine gearbox samples,considering that unlabeled data also contains useful hidden information.This paper proposes a multivariate regularized semi-supervised network model,which uses unlabeled samples to improve the accuracy of fault diagnosis for wind turbine gearbox.Firstly,the initialization network model is used to generate the pseudo label of unlabeled data,and the entropy of the pseudo label is reduced by sharpening function,and the pseudo label data set is mixed with the labeled data set;then,the consistency regularization is further introduced,and the weighted sum of the loss of the two data sets is regarded as the final loss;finally,the access of the weight attenuation of the optimizer is adjusted to make the gradient of the optimizer lower In addition,weight regularization and output feature regularization are introduced through gradient centralization to improve the generalization ability of neural network further.Finally,combined with the wind turbine gearbox’s fault diagnosis method,small sample constrained deep learning is proposed in this paper based on the above research and exploration.The parameter adjustment module is designed.At the same time,the existing problems in the research are summarized and prospected.