Research On Health Condition Monitoring Method Of Wind Turbine Based On Data-driven

In recent years,as a clean source of renewable energy with great potential,wind energy has become an indispensable part in solving world energy problems,and the global installed capacity of wind turbines has shown an upward trend year by year.However,the long-term operation of turbines will inevitably lead to failure,and the sudden shutdown caused by turbine failure will cause serious economic losses and will also have a great impact on the smooth operation of the power grid.Therefore,in order to ensure the long-term safe and reliable operation of wind turbines and reduce the downtime and operation and maintenance costs,it is very necessary to study the condition monitoring and fault detection technology of wind turbines.Based on the operating data of the Supervisory Control and Data Acquisition(SCADA)system,this thesis will research on the condition monitoring and fault detection method of wind turbines and their key components based on extreme learning machine and deep learning,which provides technical support for the realization of early fault warning and classification identification of the turbines and their components.The main work of this thesis is as follows:(1)In view of the high-dimensional redundancy characteristics of wind turbine SCADA monitoring data,a condition monitoring method of main bearing based on multivariate correlation analysis is proposed.In order to select the useful input variables and eliminate the redundancy between data,the feature selection method based on correlation coefficient method and principal component analysis method is firstly studied.Then,a SCADA data-driven monitoring model is constructed using the new neural network-extreme learning machine to capture the correlation relationship between modeling output and related input variables.The performance of the proposed condition monitoring method is verified by the fault simulation situation of the measured SCADA data,and a comparative study is carried out with the traditional monitoring method.(2)In view of the correlation coupling and high complexity between SCADA monitoring parameters and the limitation of traditional methods combing feature selection and shallow learning network,a condition monitoring method of main bearing based on deep feature learning is proposed.The deep belief network with nonlinear modeling capability is used to automatically extract features from the SCADA data to reduce the dependence on feature selection,and to deeply mine the potential fault symptoms hidden in the data,so as to realize the effective detection of the abnormal state.The prediction performance of the proposed method is verified by the measured SCADA data,and the effectiveness and advantages of the proposed method in the detection of abnormal state of the main bearing are proved by the fault simulation based on the measured data.(3)In view of the complex,changeable,and highly dynamic characteristics of wind turbine operating conditions,an early fault detection approach for wind turbines based on partial modeling under different operating conditions is proposed.First,to consider the dynamic behavior and multiple operating characteristics of turbines,an operation condition partition scheme using a clustering algorithm is proposed to partition the whole operation into multiple sub-operation conditions.Second,to overcome the shortcoming of traditional shallow structure-based methods,an optimized deep belief network modeling approach oriented to sub-operation condition interval is constructed to deal with SCADA data to capture the sophisticated mapping relationship among monitoring variables.Finally,a case study of main bearing fault detection using real SCADA data is used to validate the proposed approach,which demonstrates its effectiveness and advantages.(4)Taking into account the inherent spatio-temporal correlation characteristics of SCADA multivariable time series,based on the traditional convolutional deep belief network,a novel sensor fault detection method based on multiscale spatio-temporal correlation fusion,named multiscale spatio-temporal convolutional deep belief network(MSTCDBN)is proposed.This method aims to better capture the spatio-temporal correlation and interactive feature information hidden in the SCADA data,so as to realize the effective classification and detection of sensor faults.A generic wind turbine benchmark model is used to evaluate the proposed MSTCDBN approach,and the effectiveness of this method in sensor fault detection is verified by comparative analysis.