Bearing Fault Diagnosis Method Of Wind Turbine Based On Improved Generative Adversarial Network

As one of the most reliable and feasible renewable energy sources,wind power has been strongly supported by the government in terms of scale,technology and subsidies.China’s wind power installation reached another record high in 2021,with 15,911 new installations and 55.92 million k W of installed capacity,up 2.7% year-on-year.The accumulated installed capacity reached 170,000 units,and the installed capacity exceeded 340 million k W,an increase of 19.2 percent over the previous year.Most of the wind turbines work in coastal and desert regions where there is ample wind resources.Because of the severe operation conditions,wind turbines tend to fail.Once a wind turbine fails,it will not only cause great economic loss,but also cause the destruction of it.Therefore,it is very important to diagnose the fault of wind turbine and guarantee its safe and efficient operation.The bad operation condition of the wind turbine results in severe noise pollution in bearing fault data,which makes the sample set with a very large amount of data more complex.In addition,there are serious data imbalance problems between bearing fault data and normal data,which bring great challenges to bearing fault diagnosis.Aiming at the above problems,this paper takes Generative Adversarial Network(GAN)as the research method,In order to improve the fault diagnosis rate of wind turbine bearings,the main bearings of important transmission equipment are studied as follows:(1)Deep Convolution Generates Adversarial Networks(DCGAN)and Wasserstein Generates Adversarial Networks(WGAN),A W-DCGAN fault diagnosis model is constructed to solve the problem of data imbalance in wind turbine bearing fault diagnosis.Based on DCGAN,Wasserstein distance is introduced to solve the problems of training instability,gradient disappearance and gradient explosion in DCGAN.Bearing failure data of Case Western Reserve University were used to verify the advancement and validity of the model.The fault diagnosis rate is 97.432%.(2)The Self Attention mechanism(SA)was introduced into W-DCGAN fault diagnosis model,and the SAW-DCGAN(SAw-DCgan)fault diagnosis model was constructed.Improve the fault data generation ability of the model and speed up the training time of the model.The experimental results show that the fault classification ability of the model is more obvious after the introduction of automatic attention mechanism,and the training time is about 20% faster than that of W-DCGAN model.(3)By designing several groups of samples with different unbalance degrees,the fault visualization experiment and Wasserstein distance experiment with training cycle were carried out for the proposed model.A comparison experiment of SAW-DCGAN,W-DCGAN and DCGAN fault diagnosis methods was designed to verify that the proposed model has the advantages of strong fault classification ability,short training time and high fault diagnosis rate,and some technical support suggestions were made for the practical application of the proposed model.