AI-Based Conditional Monitoring & Predictive Maintenance for Offshore Wind Farms

Abstract

Offshore wind farm maintenance is challenged by harsh marine conditions, high operational costs, and significant safety risks associated with manual inspections. To address these challenges, a predictive maintenance framework is proposed that integrates high-frequency SCADA data with advanced artificial intelligence (AI) techniques. The framework is grounded in the key concepts of condition monitoring, anomaly detection, and remaining useful life estimation, and it is built upon theories from time-series analysis, statistical signal processing, and machine learning.

A systematic data pipeline is developed, beginning with timestamp alignment, outlier filtering, and gap interpolation to ensure data integrity. Dimensionality reduction is achieved through incremental principal component analysis, while feature selection is conducted via correlation filtering and ensemble-based importance ranking. Two complementary data representations are crafted: fixed-length multivariate sequences for recurrent and convolutional neural networks, and static feature snapshots for a gradient-boosted decision-tree ensemble. Three predictive models an LSTM-based recurrent network, a one-dimensional convolutional neural network, and a LightGBM classifier are trained and validated on subsets of the publicly available “CARE to Compare” wind turbine dataset. Temporal partitioning prevents information leakage, and class-imbalance strategies, including stratified sampling and noise-injection oversampling, are employed to enhance sensitivity to rare fault events.

Model performance is evaluated using accuracy, precision, recall, and F₁-score on held-out test segments. The deep-learning approaches are shown to capture both gradual drifts and transient spikes in sensor channels, providing early-warning windows of multiple days before predicted component failures. A real-time SCADA simulation demonstrates how staged alerts across mechanical, electrical, and environmental signals can be used to orchestrate proactive maintenance schedules, thereby reducing dependency on emergency vessel mobilizations and risky platform visits. The results indicate that sequence-aware neural networks outperform static, snapshot-based methods in fault prediction tasks, validating the critical role of temporal context in offshore condition monitoring. The thesis concludes that AI-driven predictive maintenance not only enhances turbine availability and grid reliability but also promotes crew safety and cost efficiency. Recommendations for future work include integration of environmental satellite data, deployment of edge-computing solutions for on-site inference, adoption of explainable AI techniques to build operator trust, and exploration of federated learning paradigms to generalize models across multiple wind farms without sharing raw data.