AI-based Fault Detection in Hydraulic Pitch System of Wind Turbines

Abstract

Reliable and cost-effective condition monitoring (CM) of wind turbines is essential for reducing operational and maintenance (O&M) costs and ensuring the competitiveness of wind energy. This is emphasized even more in offshore environments where maintenance interventions are particularly expensive. Among all turbine subsystems, the pitch system is the most critical, as it governs rotor aerodynamic efficiency and provides the primary safety mechanism against overspeed conditions. Failures in the pitch system can therefore cause severe downtime, high repair costs, and pose risks to turbine integrity. Despite this, the condition monitoring of pitch systems has been relatively underexplored compared to other drivetrain components.

This dissertation develops a data-driven methodology for fault detection in hydraulic pitch systems of multi-MW wind turbines using Supervisory Control and Data Acquisition (SCADA) data. Ten years of operational data from a Finnish wind farm, comprising five 2.3 MW fixed-speed turbines with hydraulic pitch systems, were analyzed together with maintenance and alarm logs. For the first time in the literature, frequent pitch system failures were systematically collected, and nine representative fault types were identified for model development.

To enable supervised learning, SCADA data were reliably annotated using a modified power curve monitoring method in combination with maintenance records. Several AI models were then trained, tested, and compared, using either key critical-characteristic features (CCFs) or the full feature set related to the pitch system. Various feature extraction techniques were also applied to improve diagnostic accuracy. In particular, an autoencoder-based feature extraction approach was applied to reduce the original forty-nine features into a compact and more informative representation of the raw signals. This compact representation was subsequently used to train diagnostic models. The developed models obtained F1-scores above 85% with the top-performing model reaching 95.5%. The best-performing model consisted of an autoencoder-based feature extractor and a support vector machines (SVM) classifier and used eight new features. Its performance offers accurate and robust fault detection under varying operating conditions.

The developed methodology provides a low-cost alternative to conventional high-frequency condition monitoring systems, as it relies solely on SCADA data while still enabling real-time fault diagnosis. Its practical benefits include reducing O&M costs, supporting predictive maintenance strategies, and extending turbine lifetime. Importantly, the developed approach is directly applicable to both onshore and offshore wind turbines with hydraulic pitch systems, where the criticality of the pitch subsystem makes reliable monitoring indispensable.