Study On Fault Diagnosis Of Wind Turbine Gearboxes Based On Deep Recurrent Network

Wind turbines often work in bad and changeable conditions such as variable wind load and large temperature difference for a long time.Under the action of alternating load,they are prone to failure,which results in units shutdown and tremendously affects their service life.Gearboxes are important part of the transmission chain of wind turbines.Their failure rate is relatively low,but it causes long downtime and high maintenance cost.Therefore,the fault diagnosis of wind turbine gearboxes is an important research topic.The traditional mode of wind power operation and maintenance is to judge the operation status of wind power gearbox according to the fault early warning information of Supervisory Control and Data Acquisition(SCADA)monitoring system,and then to locate and diagnose the fault of wind turbine gearboxes based on the vibration data of Condition Monitoring System(CMS).SCADA monitoring system often uses a single monitoring parameter to monitor the status of wind turbine gearboxes,or combines traditional machine learning methods with SCADA monitoring variables to establish a normal behavior model to monitor the status of wind turbine gearboxes.However,it is prone to misjudgement and low operational accuracy,which makes it unable to effectively monitor the status of wind turbine gearboxes.Fault diagnosis of wind turbine gearboxes based on CMS vibration signal usually adopts signal processing method or traditional machine learning method,but it needs manual intervention and unable to build complex and deep-seated nonlinear model,which ultimately lead to misjudgment and missed diagnosis of wind turbine gearboxes.Deep learning is a method which can construct deep-seated model through layer-by-layer training process,it can automatically learn discriminant features from high-dimensional data to replace traditional statistical features.The state early warning and fault diagnosis methods of wind turbine gearboxes based on deep learning have been studied in recent years.In view of the fact that both SCADA data and CMS data are time series data,this paper introduces deep recurrent network which has time series data mining ability to the study of fault diagnosis of wind turbine gearboxes.Firstly,aiming at the problem that the time series information of SCADA data can not be fully excavated by the deep learning algorithm without time memory,which leads to the low accuracy of wind turbine gearboxes state prediction,a method of wind turbine gearboxes state early warning based on long short-term memory network fusing SCADA data is proposed.In order to improve the calculation efficiency and accuracy of the model,the SCADA parameters closely related to the model condition monitoring quantity are selected as the input of the model by grey correlation analysis,and the residuals of the predicted value and the measured value is calculated to obtain the monitoring measurement for the state monitoring.The upper and lower thresholds for judging the state anomaly are calculated based on the three Sigma criterion,so as to judge the residuals change more accurately and identify it.The abnormal condition of the wind turbine gearboxes is verified by the SCADA data measured from the wind farm.Secondly,aiming at the problem that the existing fault diagnosis methods do not consider the timing information of CMS data,a fault diagnosis method of convolutional gated recurrent unit network is proposed.The original time domain features of CMS are extracted by using one-dimensional convolution neural network without destroying the temporal correlation of data.The extracted features are input into gated recurrent unit to mine the temporal information of CMS data.The left soft-saturated ELU is used as the activation function of one-dimensional convolution neural network to solve the problem that ReLU activation function unable to update some parameters of network.The validity of the proposed method is verified by the measured data from the wind farm.Finally,according to the actual needs of fault diagnosis of wind turbine gearboxes,a state early warning module based on LSTM model is developed,which can realize SCADA parameter viewing and wind turbine gearboxes state early warning;an fault diagnosis module based on CNN+GRU model is developed,which can realize CMS data waveform display in time domain and wind turbine gearboxes fault diagnosis.The effectiveness of fault diagnosis modules of wind turbine gearboxes is verified based on the actual measured data from wind farm.