Condition Monitoring And Health Evaluation Of Wind Turbines Based On Deep Learning

With the continuous increase in the capacity of my country’s wind power generation equipment and the increasing complexity of unit equipment,a series of problems such as frequent unit equipment failures,poor reliability,and high production and operation costs have become increasingly prominent,which have largely restricted my country’s wind power industry.The healthy development of the power generation industry.Wind turbine operation status monitoring and performance evaluation have gradually become the main means to solve the key problems.Based on this,this article takes wind turbine gearboxes as the main research object,and based on the deep learning method,carries out wind turbine condition monitoring and health evaluation research.The specific contents include:(1)SCADA data cleaning and preprocessing based on Change Point(CP)and Copula theory.Because wind turbines are greatly affected by various environmental conditions such as wind speed and wind direction,operating data fluctuates greatly,and are susceptible to interference,etc.,there are often many abnormalities in SCADA data,which must be cleaned and preprocessed.To this end,this paper proposes a two-stage CP-Copula algorithm for cleaning abnormal data of wind turbines.First,according to the distribution characteristics and causes of abnormal operation data of wind turbines,the abnormal data is divided into piled abnormal data and scattered abnormal data with the help of wind power curve,and the CP algorithm is used to handle a large number of piled abnormal data and a small number of scattered abnormal data to the maximum extent.data.The method uses the inflection point of the power variance rate as the basis for judging abnormal data,so as to increase the proportion of data under normal conditions.Then,based on the Copula function,the probability power curve of the correlation between wind speed and power is constructed to further clean the remaining scattered abnormal data.The algorithm is verified through the analysis of a calculation example,and the results show that the proposed CP-Copula algorithm can accurately identify different types of abnormal data,and the cleaning effect is better than other comparison methods.(2)Condition monitoring of wind turbine gearbox based on multi-parameter fusion.In view of the high-dimensional correlation,noise and uncertainty of SCADA data,this paper proposes a wind turbine gearbox state based on Batch Normalization(BN)-Sparse Denoising Autoencoder(SDAE)Monitoring model.This method first adds sparse and noise reduction dual function modules on the basis of the traditional autoencoder network to solve the problems of wind turbine key information noise submergence and weak feature extraction,and then uses the improved sparse noise reduction autoencoder network to learn the gearbox The feature rules of state data,and the introduction of batch normalization(BN)algorithm to improve network convergence speed and learning efficiency.Select the key parameter set that affects the performance of the gearbox,and use this parameter set to build a multiparameter fusion state monitoring model based on the BN-SDAE network,and then define the reconstruction error of the model,evaluate the reconstruction error,and use the kernel density estimation method Determine the performance monitoring threshold,and finally verify the proposed method with a high temperature fault of the gearbox as a case.The experimental results show the effectiveness of the algorithm.(3)Aiming at the problem that the health status of wind turbine gearboxes is difficult to evaluate,a method for evaluating the health status of gearboxes based on Long ShortTerm Memory Network(LSTM)is proposed.The method firstly uses the correlation between SCADA parameters and the evaluation parameters that reflect the state of the wind turbine gearbox as an entry point,and combines expert experience and Pearson’s correlation coefficient method to screen out the input parameters of the health state prediction model;secondly,use The most suitable LSTM network for handling time series problems predicts the selected gearbox state evaluation parameters,and calculates the residuals with the current values;according to the relative degradation of the predicted residuals,fuzzy membership and bubble charts are used to evaluate the health of the gearbox status.Finally,through the algorithm verification of the actual measured wind farm operating data,the experimental results show that the prediction algorithm and evaluation method are relatively accurate,simple and intuitive,and have certain engineering application value.