Component-based Wind Turbine Health Monitoring using Support Vector Machines on Operational and Event Data

The purpose of this thesis is to combine different condition indicators of a wind turbine to predict major failures at the gearbox component. Event and operational data from energy operators form the basis of this work. We defined and calculated different relevant condition indicators (called lifetime indicators) for two specific gearbox types. We used support vector machines (SVMs) from machine learning to analyze the data and determine when a major failure occurs. The mathematical background for SVMs was given aswell. We evaluate several SVM models on the two gearbox types. The fine-tuned SVM model for the second gearbox type yielded the following results: Out of all major failures, 48% were detected. 62% of the time when the SVM gave out a failure warning, an actual failure was going to happen within the upcoming three weeks. The "false alarms" from the SVM showed a drastically reduced mean time till the next failure. Future work could examine this SVM model to obtain insight on which lifetime indicators play a crucial role and which predictive maintenance strategies are effective in avoiding upcoming major failures.