Fault analysis of wind turbine gearbox based on long term and short term memory network

As a new machine learning method, deep learning method has been widely used in image processing, speech recognition, brain computer interface and other fields because of its powerful ability of feature learning and nonlinear system modeling. In recent years, this method has been gradually applied to the health monitoring and fault detection of wind power system. For example, deep belief network and deep neural network are proposed to model the healthy behavior of wind turbine SCADA data, and the residual between the predicted value and the real value of the model is used as the state detection quantity to realize the fault prediction of the main bearing and gearbox. However, SCADA data is essentially a multidimensional time series collected by multiple sensors. Due to the correlation coupling and interaction between different subsystems or subcomponents, there is spatial correlation between data; moreover, the current time information and historical time information of time series data have close temporal correlation. The existing methods do not consider the inherent spatiotemporal correlation of SCADA multivariable data well.

In order to solve this problem, a gearbox fault prediction method based on long short term memory (LSTM) network is proposed. Considering that the gearbox oil temperature will rise when there are various faults such as gear wear, tooth corrosion or pitting corrosion caused by fatigue, the oil temperature is selected as the target prediction variable, and the oil temperature monitoring model of normal operation of gearbox is established by fully considering the important correlation information between the oil temperature and other relevant sensor data in the space-time dimension; Furthermore, on the basis of the reconstruction residual of the normal operation model, the exponential weighted moving average (EWMA) is used to set the threshold control line for the fault prediction of the gearbox oil temperature state. Finally, an example is given to verify the effectiveness of the proposed method.

Aiming at the multivariable spatiotemporal correlation characteristics of wind turbine SCADA operation data, a fault prediction method of wind turbine gearbox based on LSTM network is proposed, aiming at deeply mining the characteristic information of SCADA data in spatiotemporal dimension, and providing a new idea for wind turbine fault prediction.

Firstly, the LSTM network model under the health condition is established to predict the oil temperature of the gearbox; then, according to the fault condition, the relationships inside the gearbox are destroyed, and the residual between the predicted value and the real value increases, so as to predict the fault. The model is compared with SVM, RR, RF and BP in two error indexes of MAPE and sdape. The results show that the proposed LSTM fault prediction model has better fitting ability to the real oil temperature data, and can build a more accurate reconstruction model. The internal structure of wind turbine is complex, and the dimension of SCADA data is large. In the follow-up research, we should pay more attention to the coupling relationship between various variables, select appropriate features, and carry out early identification and prediction of various faults.