Vibration-based imbalance detection in a population of offshore wind turbines under varying environmental conditions via neural networks and statistical time series based methods
Loading...
Date
2025-03-07
Authors
Παπαδοπούλου, Αλεξάνδρα-Αθανασία
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The increasing global demand for renewable energy has driven a rise in offshore wind turbine installations due to their higher energy yields and reduced maintenance needs compared to onshore turbines. However, offshore turbines are vulnerable to various forms of damage, including rotor imbalances, which can induce excessive vibrations, mechanical stress, and costly downtime. Early detection through Structural Health Monitoring (SHM) is crucial to mitigating these issues. This research formulates an unsupervised methodology for detecting both mass and aerodynamic rotor imbalances in an offshore wind turbine population within the same farm. The approach relies on vibrational data from nacelle-mounted accelerometers, collected over a year during normal operation when the turbine produces rated power. To enhance robustness, the methodology accounts for uncertainty factors introduced by varying environmental conditions and structural inconsistencies across turbines, ensuring a more reliable and scalable detection framework. The detection framework integrates Artificial Neural Networks (ANN) and Statistical Time Series (STS) methods. An Autoencoder (AE) is employed for ANN-based detection, while the Multiple Model (MM) concept is applied for STS-based methods. System dynamics are represented through both parametric and non-parametric techniques, including Welch-based Power Spectral Density (PSD), Cross Power Spectral Density (CSD), and parametric models such as Autoregressive (AR) and Vector Autoregressive (VAR). Vibrational signals are analyzed in three directions—side-to-side, fore-aft, and longitudinal—and two of the healthy turbines are used to train the models, establishing a reference standard for unsupervised detection capable of distinguishing between healthy and faulty turbines. The results demonstrate that the detection methods generally provide reliable performance. Both non-parametric (PSD and CSD) and parametric (AR) modeling effectively represent system dynamics for imbalance detection, except for VAR modeling. MM-based methods achieve the highest overall detection rates for both types of imbalances; however, AE-PSD outperforms all other approaches in mass imbalance detection. The fore-aft and longitudinal axial signals are more effective for detecting aerodynamic imbalance, while the side-to-side signals are better for mass imbalance, aligning with the distinct vibration patterns of each imbalance type. These findings highlight the effectiveness and scalability of the proposed methodology for detecting rotor imbalances across multiple turbines within a wind farm under varying environmental conditions.
Description
Keywords
Offshore wind turbine population, Rotor imbalance detection, Neural networks, Statistical time series method, Acceleration data, Varying environmental conditions