Research On Anomaly Identification Method Of Wind Turbine Generator Based On Convolution Neural Network(CNN)

With the continuous growth of global wind power capacity and the rapid expansion of wind farms,the minor faults in the maintenance of wind turbines may bring about expensive maintenance costs.In order to reduce WT failures and the economic losses caused by faults,condition monitoring and fault diagnosis of wind turbines are playing an increasingly important role in reducing downtime and maintenance costs.As a necessary condition for fault diagnosis,the research of abnormal identification is of great significance.At present,there are small anomalies in the field of abnormal identification of wind turbines,and many kinds of complex anomalies are difficult to identify.In order to solve these problems,this thesis presents a wind power generator anomaly identification method based on convolution neural network(CNN).In this thesis,the SCADA data collected from the wind farm are researched and analyzed.And the following four tasks are completed based on the SCADA data analysis:Finishing the SCADA data preprocessing;Finishing the clustering of wind turbine attributes;Realizing the class dimensionality reduction of fan attributes;Realizeing the identification of wind turbines based on convolution neural network(CNN).Aiming at the problem of SCADA data collected on site contains a large amount of noise and failure data,the SCADA data is preprocessed by the rules of design data cleaning and wavelet threshold denoising.After the failure data and the abnormal non-related attributes are cleaned,the retained fan attributes are denoised by the wavelet threshold denoising to obtain the denoising data.Aiming at the problem of insufficient data redundancy and incomplete feature information,this paper proposes a method to deal with attribute reduction after clustering.Based on the kmeans clustering,the fan properties are normalized to the pretreated SCADA data.After the maximum number of clusters is determined by the criteria,the clustering effect is referenced to determine the clustering result by using the contour coefficients.Finally,the attributes are divided into seven categories,and the way of grayscale presentation is put forward to verify the superiority of the clustering effect.The dimension of clustered attributes are reduced based on t-SNE and the important attributes of the fan attributes are represented with the most simplified features.Aiming at the problem that the traditional model is prone to misjudgment in the process of anomaly identification,this paper proposes an anomaly identification method based on CNN.The input attribute value data is transformed into square images of many fan attributes of the same size.Through training mass images,the model parameters are continuously adjusted to obtain the structure of each layer of the CNN model,and the judgment of fan status is realized through the dichotomy of the pictures.If the difference between the data caused by the weather factor and the data caused by the fault is small,the CNN model can restrain the unfavorable interference of noise and accurately find the slight change of the picture of the wind turbine property,so that the state of the wind turbine can be included in the normal category and the mistake judgment.Different from the traditional modeling based on a certain attribute,we can only find out the limitation of the abnormality of a single attribute.The input sample of CNN model is provided by a complete WT image segment,which contains comprehensive exception information and can effectively realize the abnormality of multiple attributes recognize.Aiming at the method proposed in this paper,three examples are given in the experimental part to prove the feasibility and effectiveness of the proposed method.Finally,the future research direction is prospected based on the summary of the full text.