Evaluation and Application of Transformer Architectures for Image-Based Detection and Localization of Damage to Wind Turbines

Due to their location in open areas, wind turbines (WT) are particularly susceptible to environmental damage, which can affect energy production and profitability. Automated damage detection systems based on deep learning offer a way of identifying and localizing damage at an early stage. Transformer architectures, which are becoming increasingly important in computer vision, have not yet been used in damage localization. This work compares the transformer architectures ViTDet and Swin Transformer with the Convolutional Neural Network (CNN) architecture ResNet in terms of their performance, robustness and computational complexity in damage localization. Two datasets were used for the analysis: a publicly available dataset and a private dataset that contains thermal images in addition to RGB images. It was also analyzed whether the use of thermal images together with RGB images offers additional benefits for damage localization. The results showed that transformer architectures, in particular the Swin Transformer, are superior to CNN-based approaches in terms of performance. No significant differences were found in terms of robustness to weather, perspective and motion blur perturbations. However, the transformer architectures are less efficient than the CNNs, with the Swin Transformer performing better than ViTDet. Thermal images led to a minimal improvement in the performance metrics, but also increased the model complexity by using both modalities in parallel.