A transfer learning approach for mitigating temperature effects on wind turbine blades damage diagnosis
Nima Rezazadeh, Fawaz Annaz, Waheb A. Jabbar, Jozué Vieira Filho, Mario De Oliveira
Abstract
Data scarcity, coupled with environmental and operational variabilities (EOVs), poses substantial challenges to the generalisability and robustness of damage diagnostic methods for complex components such as wind turbine blades. This paper introduces a novel methodology, termed UCTRF, designed to tackle these challenges. UCTRF stands for Uniform manifold approximation and projection for dimensionality reduction, Capsule neural networks for advanced feature recognition, Transfer adaptive boosting for effective knowledge transfer, and Random Forest for nuanced instance weighting and classification. The UCTRF framework is uniquely suited to scenarios where feature distributions shift due to temperature variations, enabling robust knowledge transfer even in limited datasets. This innovative framework was rigorously evaluated on various temperature-affected datasets, achieving a 95% detection rate. These results underscore its effectiveness in preserving the structural integrity of wind turbines under challenging EOVs and constrained data availability. Additionally, the internal mechanism of the designed domain adaptation captures the alterations in instance weights between the source and target domains during the adjustment process, which can be utilised to analyse the impact of diverse instances on model performance and further refine the adaptation process.