Data Expansion and Defect Classification Method for Small Aero-Engine Blade Samples: Low-Rank Adaptation and Edge Information Enhancement

Deep learning-based defect classification of aero-engine blades faces two critical challenges: scarce training samples and domain shift between laboratory-collected training data and real-world production environments. This study addresses these issues through three main contributions: (1) We introduce aero-engine blade dataset with domain shift (AeBDDS), a novel cross-domain benchmark specifically designed to evaluate model robustness under domain shift; (2) we employ low-rank adaption (LoRA) to fine-tune the stable diffusion for generating high-fidelity synthetic samples that approximate production conditions, effectively augmenting the training data; (3) we propose edge information enhancement and multiple attention network (EIEMANet), a classification framework that leverages edge information enhancement and multi-scale attention mechanisms to capture contextual and damage-specific features. Experiments demonstrate that training with mixed synthetic and real data significantly improves cross-domain generalization. Under the T&S_4-6 setting, EIEMANet achieves over 86% in accuracy, precision, recall, and F1-score, demonstrating clear advantages over baseline methods in domain-shift scenarios.