Assessment of Vibration-based Fault Diagnosis in Offshore Wind Turbines under Real-World Operating Conditions

Thesis Abstract

This study addresses the critical challenge of maintaining reliability and reducing downtime in offshore wind turbines by advancing vibration-based fault diagnosis under operationally realistic conditions. The aim is to develop a robust, data-driven diagnostic framework capable of identifying incipient faults in drivetrain and rotor components using vibration signals recorded during regular and extreme weather conditions. Specific objectives include (1) to characterize normal operational vibration signatures for key subassemblies (gearbox, bearings, generator, and blades) across a 24-month offshore dataset; (2) to detect and classify fault conditions at early stages through feature extraction and selection, employing time-domain, frequency-domain, and time–frequency analyses; (3) to establish a validated diagnostic model integrating machine learning with physics-based priors to improve fault localization accuracy; (4) to assess the model’s robustness to environmental and operational variability such as wind speed, surge, and wave-induced loads; and (5) to translate diagnostic outcomes into actionable maintenance recommendations and risk-informed decision criteria. The study adopts a mixed-methods approach, combining empirical vibration data with expert maintenance records to triangulate fault labels. The population comprises offshore wind turbines installed in a representative offshore wind farm cluster with heterogeneous turbine models. A stratified random sample of 60 turbines is selected, capturing variations in turbine age (3–12 years), gearbox configurations, hub heights, and environmental exposure. Data collection relies on high-frequency accelerometers (?12 kHz sampling) mounted near bearings, gears, and nacelle mounts, complemented by enriched metadata from SCADA systems, maintenance logs, and meteorological measurements. The analysis employs a sequence of methods (i) signal preprocessing and artifact removal; (ii) feature engineering including spectral kurtosis, wavelet packet energy, empirical mode decomposition components, and statistical moments; (iii) feature selection through mutual information and recursive feature elimination; (iv) supervised learning using random forest, support vector machines, and gradient boosting, coupled with Bayesian network constraints to incorporate physics-based priors; (v) model validation via nested cross-validation, with performance metrics including precision, recall, F1-score, and area under the ROC curve; and (vi) sensitivity analysis to quantify robustness against environmental variability. Ground-truth fault labels are established from maintenance records corroborated by vibration-based indicators and expert assessments, while a subset of 12 turbines undergo controlled fault injection simulations in a validated digital twin environment to further test diagnostic resilience. The theoretical framing integrates fault-diagnosis theory with reliability-centered maintenance and data-driven modeling, drawing on the Health Monitoring and Diagnostics Theory and the Bayesian Updating framework to fuse prior knowledge with observed data. Expected findings include (a) discriminative feature sets capable of early fault detection (incipient bearing wear, gear tooth damage, misalignment) with rising diagnostic accuracy as data accumulate over time; (b) a generalizable diagnostic model robust to wind and wave-induced noise, validated across turbine models; and (c) quantified thresholds enabling condition-based maintenance triggers that reduce unplanned outages by an estimated 15–25% and extend mean time between failures. The study contributes to knowledge by (i) delivering a validated, scalable vibration-based diagnostic framework tailored for real-world offshore environments, (ii) advancing the integration of physics-informed machine learning with offshore condition monitoring, and (iii) providing transferable guidelines for implementing proactive maintenance in offshore wind portfolios. The main conclusion posits that a hybrid model combining data-driven classifiers with physics-based priors can reliably detect and localize faults in offshore wind turbines under diverse operating conditions, enabling timely maintenance and improved asset availability. Recommendations include deploying the framework across additional offshore sites with expanded sensor networks, refining digital twin simulations to cover multi-age fleets, and embedding the diagnostic outputs within operator maintenance decision-support dashboards to optimize lifecycle management.