| Compared with equipment working in a relatively stable condition,such as steam,hydraulic or gas turbines in traditional power plants,wind turbines are usually located in a harsher environment,resulting in a higher failure rate.In order to reduce the impact of power plants on the ecological environment,the world’s growing demand for wind energy has led to a continuously increase of the size and rated power of wind turbines in recent years which reduces their reliability further.Therefore,continuous condition monitoring of wind turbines and early fault diagnosis is conducive to reducing operation and maintenance costs,improving repair efficiency,and avoiding shutdown or even catastrophic events due to structure damage.With the support of sensor monitoring and big data analysis technology,multi-modal information reflecting the wind turbine’s continuously changing performance,such as voltage,current,power and temperature,can be completely documented by the supervisory control and data acquisition(SCADA)system at a low cost,providing sufficient data support for condition monitoring and fault diagnosis of the turbine.However,SCADA data are usually time-varying,coupling,and fail to include all the fault samples.Therefore,how to deeply extract fault features from existing SCADA data for early fault identification is of great significance for formulating maintenance strategies scientifically and rationally.On these grounds,the specific research work of this paper is conducted as follows:(1)In view of the time-varying and cross-coupling characteristic of SCADA data,a wind turbine condition monitoring method based on Long Short-Term Memory-Auto Encoder(LSTM-AE)neural network is proposed.Firstly,according to the distribution characteristics of the missing values and noise values in the original data set,list deletion method and multiple classification model algorithm is used separately in data preprocessing to reduce model instability caused by data quality.Secondly,the hidden layer neurons of AE are replaced by LSTM,so that the model is capable of extracting temporal features and dealing with nonlinear relations simultaneously.Finally,considering the model evaluation deviation due to environmental factors,an adaptive threshold method based on support vector regression instead of fixed threshold is presented to monitor the changes of real-time wind turbine operating conditions.(2)In response to the problem of incomplete fault data,a mutual information estimation method based on Copula entropy is proposed to quantify the correlation between SCADA features and abnormal condition obtained from the results of condition monitoring model,which is defined as the condition when the wind turbine operating condition fails to meet normal requirements and has not yet been shut down.Highly correlated features are then selected as fault features and used to build an improved fault diagnosis framework based on gaussian mixed model.Finally,a case study of the wind turbine near a coast of Ireland is employed to verify the effectiveness of the proposed method.The results show that the SCADA data-based wind turbine anomaly detection and fault diagnosis method proposed in this paper improves model accuracy,and provides an effective solution for previous evaluation model drawbacks of lacking consideration on SCADA data’s time-varying characteristic and low accuracy subject to the fixed threshold.Meanwhile,not only can the proposed method identify new fault categories that are not included in the training set effectively,but also can provide probabilistic results to assist engineers in generating maintenance strategy. |
