04.2 - Aerostructures Design, Structural Dynamics, Aeroelasticity
SEMI-SUPERVISED MULTI-TASK CONSISTENCY LEARNING FOR DEFECT DETECTION AND DEPTH ESTIMATION
S. Miao¹, Y. Ge¹, Y. Pei¹; ¹Northwestern Polytechnical University, China
Damage tolerance and high survivability of aircraft structures are key design indices for modern aircraft. When a wing box sustains composite damage, the elastoplastic coupling mechanism significantly reduces its load-bearing capacity. Detection, classification, and quantitative assessment of composite damage are critical for evaluating structural load-bearing capacity. However, existing methods often face challenges such as difficulties in multi-task collaboration and scarcity of labeled data. To achieve efficient multi-task synergy from detection to quantitative analysis, this paper proposes a consistency regularization-based semi-supervised multi-task learning framework (MTCL). The framework deeply integrates damage detection/classification and dent depth estimation through a multi-task learning mechanism, utilizing YOLOv8 as the backbone for shared feature extraction. Specifically, one branch focuses on the localization and classification of multiple types of damage, while the other branch performs pixel-level depth estimation for dent defects. In addition, to address the challenge of scarce real labels in multi-task learning models, a Mean Teacher collaborative training mechanism is proposed. This mechanism combines supervised loss with unsupervised loss based on consistency regularization, forming a joint optimization objective that enables the fusion of labeled and unlabeled data. For the dent depth estimation task, a scale-aware depth regression module is designed, which decouples relative depth prediction from absolute scale estimation and achieves precise quantification of pit defect depth with only a small amount of real depth-annotated data. Validation demonstrates that the framework exhibits superior data utilization efficiency under conditions of scarce label data. It not only provides a technical pathway for quantitative assessment of composite structure damage, but also offers a referential framework for visual inspection and depth estimation through
DOI: 10.71945/icas2026_0072