Research on Unsupervised Fault Diagnosis Method for Wind Turbine Gearbox

Aiming at strong nonlinear characteristics of wind turbine gearbox vibration signals, an improved variational mode decomposition method was proposed to decompose signals for extracting characteristic components, and the nonlinear changes of the signal were quantified by chaotic phase portraits and Lyapunov exponent. To ensure the reliability of fault feature extraction and improve the accuracy of fault diagnosis, the random nearest neighbor embedding algorithm was used to reduce redundant features of multi-modal nonlinear fault feature sets. The proposed method was applied to NREL GRC wind turbine gearbox faults due to the unsupervised fault diagnosis framework being more suitable for engineering applications without manual marking of fault samples. Results show that the improved variational mode decomposition method can accurately extract multi-modal features. Combined with the random nearest neighbor embedding algorithm, redundant features can be effectively eliminated to ensure the reliability of fault information. Moreover, the clustering of similar samples and the difference of heterogeneous samples increase, and the clustering performance is clearer, which improves the accuracy of fault classification.