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

Area of Science:

• Artificial Intelligence
• Mechanical Engineering
• Materials Science

Background:

• Deep learning for aero-engine blade defect classification is hindered by scarce training data and domain shift.
• Real-world production environments differ significantly from laboratory data, impacting model performance.

Purpose of the Study:

• To develop a robust deep learning framework for aero-engine blade defect classification addressing data scarcity and domain shift.
• To introduce a new cross-domain benchmark dataset (AeBDDS) for evaluating model robustness.

Main Methods:

• Utilized low-rank adaptation (LoRA) to fine-tune stable diffusion for generating high-fidelity synthetic training samples.
• Proposed the Edge Information Enhancement and Multiple Attention Network (EIEMANet) for feature extraction.
• Integrated synthetic and real-world data for comprehensive model training.

Main Results:

• EIEMANet achieved over 86% accuracy, precision, recall, and F1-score under the T&S4-6 domain-shift setting.
• Training with mixed synthetic and real data demonstrated significant improvements in cross-domain generalization.
• The proposed methods outperformed baseline approaches in domain-shift scenarios.

Conclusions:

• The developed framework effectively addresses data scarcity and domain shift challenges in aero-engine blade defect classification.
• High-fidelity synthetic data generation and advanced attention mechanisms enhance model robustness and performance.
• The AeBDDS benchmark facilitates future research on domain generalization for critical industrial applications.