Data-Driven Fault Warning Method For Wind Turbine Transmission Components

Under the policy of "30-60 Double Carbon Target",the wind power industry in China will enter a high-speed development stage.Due to the complex internal structure and harsh operating environment of wind turbines,frequent failures of wind turbines have resulted in decreased power generation and increased operating and maintenance costs.Potential faults are found in advance and eliminated in time through advanced automated fault warning technology,which can effectively save repair costs and reduce downtime.Therefore,this paper takes SCADA data with low cost,rich sensor types,and long recording time as the data source to study the fault warning method based on normal behavior modeling.The implementation process includes data processing,feature selection,data-driven modeling and residual analysis.In this paper,each part is studied in depth,and the corresponding methods are proposed.The research content is as follows:In chapter 2,an anomaly data processing method based on density clustering and boundary extraction is proposed.The data processing phase is to remove null and anomalous values from the raw data,avoiding the negative effects of anomalous data on data-driven modeling.Anomaly data is generally classified into three categories:bottom stacked anomaly(Type 1),power limited anomaly(Type 2),and scattered anomaly(Type 3).According to the distribution characteristics and processing difficulties of Type 1,combined with the operation characteristics of wind turbines at low wind speeds,an intuitive rule method was designed to remove the Type 1.According to the distribution characteristics and processing difficulties of Type 2,a processing method combining boundary extraction and boundary trimming was designed to remove the Type 2.According to the distribution characteristics and processing difficulties of Type 3,the density peak clustering algorithm was improved and used to remove the Type 3.In order to avoid the interaction of all kinds of anomaly points,partition clustering is added to the front end of the processing flow to improve the overall processing effect.In chapter 3,a feature variable selection method based on neighbour component analysis is proposed.Feature selection is to select key feature variables from a large number of parameter variables that play an important role in data-driven modeling,reducing data dimensions and model complexity.In view of the problem that the existing feature selection methods are not applicable to the normal behavior modeling of wind turbines,this chapter proposes a feature variable selection method based on the neighbor regression algorithm for the fault warning of wind turbines based on normal behavior modeling.The proposed method determines the importance weight by calculating the contribution rate of each variable in the regression modeling process.By analyzing the characteristics of redundant variables in SCADA data,the redundant variables are removed by using correlation coefficient matrix.Experiments show that the feature selection results of the proposed method are more intuitive,can remove redundant variables,and is helpful to improve the performance of fault warning.Data-driven modeling is to train the normal behavior model of the monitored component by using the historical data of normal state.Residual analysis is to use the residual between the measured and predicted values of the target variable to calculate the fault warning index for monitoring the condition of the equipment.The normal behavior models used for wind turbine fault warning are mainly divided into two categories:variate prediction models and multivariate reconstruction models.The application scenarios of the two models are different,so this paper studies the fault warning methods based on the two models respectively.In chapter 4,based on data processing and feature selection,the data-driven modeling and residual analysis of univariate prediction model are studied.In order to study the temporal and spatial features of the data,the autocorrelation coefficient and partial autocorrelation coefficient are used to analyze the temporal features,which verify the existence of multi-scale short-time correlation.Combined with the working principle of the oil cooling system,the relationship between the long-term time features and the operation process is analyzed,and the existence of the long-term time features is verified.The existence of multi-scale spatial features is verified by using the maximum information coefficient to analyze spatial features.In view of the above spatial-temporal characteristics,a multi-scale short-time feature module is constructed by using the echo state network,a long-term feature module is constructed by using the gate recurrent unit network,and a multi-scale spatial feature module is constructed by using the dilated convolution network.Combining the three modules,a multi-scale spatial-temporal depth network model is proposed.Based on the single variable predicted residuals,residual control chart and relative entropy are used as residual analysis methods to achieve fault warning.In chapter 5,based on data processing and feature variable selection,the data-driven modeling and residual analysis of multivariable reconstruction model are studied.Aiming at the shortcomings of the existing multivariable reconstruction models,such as poor learning ability of spatio-temporal features and difficulty in fully mining the deep information of complex SCADA data,a BiLSTM-VAE multivariable reconstruction model is proposed by combining bidirectional long short-term memory neural network(BiLSTM)and variational auto encoder(VAE).The proposed model can learn spatial-temporal features and has better reconstruction performance.After the multivariate reconstruction residual is obtained,it is too heavy to analyze each residual variable one by one,so it is necessary to further use the multivariate residual analysis method to calculate a fault early index.However,the existing multivariable residual analysis method is difficult to highlight the anomalous trend effectively.To solve this problem,an index calculation method based on variable grouped Mahalanobis distance is proposed.After the implementation of fault warning,fault isolation is needed in order to determine the fault variables.Aiming at the problem that the existing methods do not consider the coupling effect between residual variables,a fault isolation method based on random forest was proposed.