Título: Comparing performances of deep learning models to classify and detect defects in wind turbine blades
Autores: Leticia Vieira Gonçalves, Paulo Henrique Araújo Bezerra & Rosana C. B. Rego
Resumo: Automated inspection has become a relevant tool for structural health monitoring in civil infrastructure, industrial, and energy generation domains, especially in components subject to harsh environmental conditions. This study presents a comparative evaluation of three Deep Learning (DL) architectures for detecting surface cracks in wind turbine blades: a sequential Convolutional Neural Network (CNN) for image classification; a U-Net architecture for segmentation, combined with a CNN classifier; and an object detection model based on an adapted YOLOv8 architecture. All models were trained and evaluated using a public dataset, which comprises images of wind turbine blades subdivided into healthy and faulty classes. Oversampling was employed to improve data quality. Each model was assessed based on performance metrics and suitability for crack detection in blade images. The CNN model provided a baseline for classification tasks, while the U-Net model demonstrated enhanced sensitivity to fine discontinuities. The YOLOv8-based detector allowed simultaneous classification and localization of defects. This study highlights trade-offs between classification and detection approaches, emphasizing the importance of selecting appropriate models depending on application constraints.
Palavras-chave: Deep Learning; Defect; Detection; Wind Turbine; Inspection.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1191827
Artigo em PDF: CBIC_2025_paper1191827.pdf
Arquivo BibTeX:
CBIC_2025_1191827.bib