A multi-target damage MS-SEG detection method for wind turbine blades surfaces equipped with errors and missed detection corrections

Abstract

Abstract Highly accurate segmentation and location of blade surface damage are essential for its criticality assessment. However, small- and medium-sized target damage on the blade surface is likely to be erroneously and not detected because of complex natural scenes and blade surface texture. To address this issue, an MS-SEG method–fused MaskCut unsupervised segmentation and Segment Anything Model (SAM) error detection correction is established to segment and localize multitarget damage on the blade surfaces individually. First, the MaskCut method is adopted to perform unsupervised individual segmentation of multitarget damage on the blade surface and construct multiple single-target damage masks. The SAM is then incorporated to interactively correct for missed and erroneous detection impairments in MaskCut unsupervised segmentation. Finally, marking anchor boxes for positioning multitarget damage on an individual basis follows the strong separability of contour and noncontour pixels. Numerical and experimental studies on wind turbine blade surfaces are conducted to validate the effectiveness of the proposed model. The results show a high segmentation and location accuracy of multitarget damage. Meanwhile, the Pixel Accuracy (Pa), Intersection of Union (IOU), and F1-score values are increased by 30.74%, 55.45%, and 49.28%, respectively, concerning the precorrection. The average accuracy of Pa, IOU, and F1-score for nine types of damage detection are 97.75%, 86.49%, and 92.48%, respectively, which show strong robustness to the detection of multiple damage features.