Multi-order nonstationarity-driven cointegration method for wind turbine generator condition monitoring and anomaly detection

Wind turbine Supervisory Control and Data Acquisition (SCADA) data are inherently nonstationary due to changing environmental conditions, operating regimes, and component aging, while emerging faults can further alter the statistical stationarity behavior of monitoring parameters. These characteristics make reliable condition monitoring and fault detection particularly challenging. This paper proposes a novel multi-order nonstationarity-driven cointegration framework for wind turbine monitoring that explicitly accounts for variables with different levels of nonstationarity. Using Johansen’s cointegration method, long-term equilibrium relationships are identified from the training data, and normalized cointegrating vectors are estimated. Cointegration residuals are then obtained by projecting SCADA data onto these vectors. Stationary residuals from the training set are used to establish control limits, while monitoring statistics, based on the Mahalanobis distance, are calculated using stationary residuals from the testing set. An alarm is triggered when these statistics exceed the defined threshold. The proposed approach is validated through a numerical study involving simulated faults in time series with different nonstationarity levels, followed by application to real SCADA data from a commercial wind turbine. Results demonstrate that the method effectively handles multi-order nonstationarity and enables accurate early detection of generator faults. The proposed framework detected the faults approximately 40 min and 33 h and 7 min earlier than the actual occurrence of generator bearing damage on August 20, 2017, at 06:08 and generator damage on August 21, 2017, at 14:47. The early detection capability highlights its potential to improve the robustness and reliability of wind turbine condition monitoring systems compared to principal component analysis and partial least squares.