A Research And Application For Wind Turbine Health Condition Monitoring Based On Data-driven

As wind turbines has provided a certain amount of electricity in China in the past few years,China becomes the first country with installed wind farm power worldwide.Operation and maintenance(O&M)including insurance,regular maintenance,repair,spare parts,and administration costs constitute a sizable share of the annual costs of wind turbine.Specifically,the key parts,in wind turbine system,are often exposed to extreme environmental variables,such as frost and typhoon weather conditions.Such conditions make the core parts prone to frequent damages and various kinds of failures.Thus,O&M costs are attracting greater attention.Manufacturers attempt to lower these costs significantly by developing new algorithms,and these algorithms require fewer regular service visits and less turbine downtime.To this end,Prognostics and Health Monitoring(PHM)is vital for wind turbine system.Unfortunately,most of current PHM approaches suffered some intractable challenges such as unbalanced data,poor domain adaption ability,and these approaches are obviously dependent on a rich understanding of wind turbine system physical model and human expertises.In order to overcome these aboved limitations as well as avoid catastrophic accidents and reduce O & M costs,a new intelligent PHM scheme is proposed and it can prevent failure by providing early alerts as well as enable better maintenance decision.The main content can be summarized as fellows.To detect abnormal condition under healthy data,an unsupervised method based on deep constitutional autoencoder(DCAE)is designed.In training process,the health samples are used to feed to DCAE model.Then,the latent features can be learned by this new model.In test process,an index created by reconstruction error is applied to identify the corresponding abnormally conditions of wind turbine blade.The efficiency and feasibility of DCAE model is demonstrated by real-world dataset.To adapt data imbalanced problem in wind turbine system,a new method based on conditional generative adversarial networks(CGAN)is introduced.In this new algorithm,a core step for data augmentation is applied to obtain amounts of failure data sequence.Then,an open access bearings dataset from Case Western Reserve University is used to improve the ability of classification.To track the long-term condition forecasting of wind turbine bearings,a fusion algorithm between long short time memory(LSTM)network and generative adversarial network(GAN)is proposed.In this method,the Nash-Equilibrium is created by discriminator output in GAN and it can consider as a health index for wind turbine bearings health condition monitoring.Moreover,the LSTM network has great advantage in forecasting condition.Hence,this fusion algorithm can precisely forecast the degradation curve in bearings compare with traditional neural networks.