YOLO-CBAM: Arquitetura YOLOv8 leve com módulo de atenção para segmentação de defeitos em pás de aerogeradores

Abstract

The increasing global demand for clean and sustainable energy has intensified the adoption of renewable technologies, with wind power standing out as one of the most promising sources. Ensuring the structural integrity and operational efficiency of wind turbines is essential, as their blades are constantly exposed to harsh environmental conditions that cause erosion, cracks, and surface wear. Conventional visual inspections, however, remain limited by human subjectivity, high operational costs, and low scalability. In response to these limitations, this work presents an automated visual inspection approach based on deep learning, proposing a lightweight segmentation model derived from the YOLOv8 architecture and enhanced with the Convolutional Block Attention Module (CBAM). The inclusion of attention mechanisms enables the network to better capture spatial and channel-wise features, improving its ability to detect subtle and low-contrast defects while maintaining computational efficiency for real-world deployment. The proposed model, named YOLOv8+CBAM, was trained on cropped and annotated samples from the public Blade30 dataset. Experimental results demonstrate significant performance improvements, with the mean Intersection over Union (mIoU) increasing from 0.57 to 0.61 compared to the baseline YOLOv8. The method shows high sensitivity to small defects, robustness against class imbalance, and efficient inference suitable for resource-constrained environments. Overall, this study contributes to advancing automated defect detection in wind turbine blades by integrating UAV-based imaging and deep learning techniques into a unified segmentation pipeline. Future extensions may explore defect severity quantification and integration with image mosaicking systems to support fully autonomous inspection and maintenance workflows.