Optimizing Wind Turbine Gearbox Diagnostics in Offshore Farms

Thesis Abstract

The reliability and availability of offshore wind turbine gearboxes are critical determinants of farm-level energy yield and maintenance costs, yet current diagnostic practices struggle with environmental harshness, limited access, and data fragmentation across multiple turbine models. This study addresses the gap by developing an integrative diagnostic framework that enhances fault detection, remaining useful life estimation, and maintenance decision support for offshore gearboxes. The aims are to (1) improve early fault detection through multimodal condition monitoring, (2) refine remaining useful life (RUL) predictions using data-driven models, and (3) provide a decision-support tool that optimizes maintenance scheduling under logistics and safety constraints. Specific objectives include (a) compiling a comprehensive dataset from a coastal offshore wind farm consisting of 72 turbines of two major gearbox platforms, with 24 months of operation data, (b) evaluating vibration, oil analysis, temperature, and SCADA signals for fault indicators, (c) developing a hybrid diagnostic model combining machine learning with physics-based features to detect gear and bearing faults, (d) comparing regression-based RUL models (Gaussian process, random forest, and neural networks) against a physics-informed prognostic model, and (e) validating the framework against historical maintenance records to quantify reductions in unplanned outages and maintenance costs. A robust, mixed-methods approach is employed. The population comprises offshore wind turbines operated by a leading European utility, with a stratified sample of 60 turbines representative of both gearbox platforms. Data collection involves sensor streams from condition monitoring systems (vibration, oil debris, oil fertility, gear temperatures), SCADA data, and maintenance logs, totaling over 1.2 terabytes of structured and unstructured information. Instruments include calibrated vibration accelerometers, oil quality analyzers, and standardized maintenance record templates. The study adopts a sequential explanatory design, wherein quantitative analyses inform subsequent qualitative validation with expert interviews (n=15 maintenance engineers) to contextualize model outputs. Validity and reliability are ensured through cross-validation, time-series holdout testing, and tracer-based data quality checks, with imputation for missing sensor readings using multiple imputation by chained equations. Variable selection relies on mutual information and recursive feature elimination, while model development uses a hybrid framework (i) signal processing with wavelet packet decomposition and spectral kurtosis to extract fault-sensitive features, (ii) a physics-informed neural network (PINN) that integrates bearing and gearbox fault physics, and (iii) ensemble learning (stacked generalization) combining Gaussian process regression for uncertainty quantification with neural networks for nonlinear pattern capture. Model performance is evaluated using accuracy, F1-score for fault classification, RMSE for RUL predictions, and reliability metrics such as time-to-failure calibration curves. Statistical comparisons utilize ANOVA and pairwise t-tests to assess significant improvements over baseline diagnostics. The ethical considerations incorporate data governance, instrument calibration standards, and safety protocols for offshore data collection. Expected findings indicate that the multimodal diagnostic framework outperforms conventional vibration-only approaches, with a 25–35% improvement in early fault detection latency and a 15–25% reduction in unplanned outages over a 12-month validation period. RUL predictions are anticipated to achieve a mean absolute percentage error below 12% with credible intervals well-calibrated through Gaussian process components. The qualitative interviews are expected to reveal organizational and logistical factors that influence maintenance decision-making, informing the design of the decision-support tool. The study contributes to knowledge by demonstrating the value of integrating physics-based constraints with data-driven models in offshore wind gearbox diagnostics, extending prognostic methodologies to multi-source condition monitoring in maritime environments, and providing a replicable framework for utility-scale deployment. The main conclusion posits that an integrated, uncertainty-aware diagnostic framework substantially enhances gearbox health management in offshore farms, enabling more reliable energy production and cost-effective maintenance planning. Recommendations include adopting the framework as a standard operational tool across offshore wind portfolios, investing in data standardization across turbine models to facilitate cross-site transferability, and developing fallback maintenance strategies under adverse weather windows to minimize downtime. Potential avenues for further research involve extending the approach to multi-asset fleet optimization, incorporating satellite-based environmental data to contextualize corrosion and wear, and exploring real-time edge-computing implementations for on-site diagnostics with reduced data bandwidth requirements.