Prediction of Gearbox Condition Changes in Wind Turbines : Anomaly Detection in Temperature and Vibration Data Based a Long Short Term Memory (LSTM) Model

Sammanfattning

Detection of early condition changes in wind turbine gearboxes is of great interest, as it leads to reduced downtime and maintenance-related costs. The aim of this study is to address the gap between feature-based Long Short Term Memory (LSTM) machine learning models and residual based anomaly detection to identify abnormal behavior in temperature and vibration data from wind turbine gearboxes from Skellefteå Kraft. This study is limited to investigate bearing faults. Specifically, it answers the two research questions regarding how residual based anomaly detection can be applied to detect early condition changes in turbine gearboxes using oil temperature and vibration data. Furthermore, the study investigates how a supervised LSTM model can be used to identify patterns related to early signs of condition changes and bearing faults. The methodology to achieve this goal consists of separate data preprocessing of temperature and vibration data, since they differ in the number of observations. Each dataset is divided into training data from turbines under normal operation, test data from one normally operating turbine, and test data from a faulty turbine with a known bearing fault. The training data is processed through binning, Inter Quartile Range (IQR) filtering, and normalizing/scaling before being used for LSTM model training. During the training phase, the dataset is divided into 80% training data and 20% test data. After training, the normal operation and faulty test turbines are plotted individually with the predicted values and ground truth data to enable comparisons. Furthermore, the rolling mean is applied to the residuals between ground truth and predicted values for the purpose of detecting abnormal behavior. The results illustrate how the temperature data and vibration data produce the anomaly detection results for identifying abnormal behavior, although the results should be interpreted as indicators rather than definitive evidence of faults.