Generative Feature Enhancement and Metric Learning for Few-Shot Defects in Aero-Engine Blades

Abstract

The structural integrity of aero-engine blades is paramount to aviation safety, requiring rigorous inspection protocols to identify surface and sub-surface defects. While deep learning has revolutionized industrial defect detection, its efficacy is severely constrained in aerospace applications by the paucity of defect samples, a condition known as the few-shot learning problem. Standard convolutional neural networks struggle to generalize from limited data, leading to overfitting and poor recognition of rare defect classes. This paper proposes a novel framework that integrates Generative Feature Enhancement with Metric Learning to address the scarcity of labeled defect data in aero-engine blades. Unlike traditional data augmentation techniques that operate in the image space, our approach synthesizes diverse training samples in the high-dimensional feature space using a modified generative adversarial network. Concurrently, a metric learning module optimizes the embedding space to ensure that defects of the same class are clustered tightly while maintaining significant separation from non-defect samples and other defect categories. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art few-shot learning algorithms in identifying micro-cracks, erosion, and ablation on blade surfaces. The dual-branch architecture ensures robust generalization capabilities even when training examples are reduced to single digits per category.