The dataset generation process begins with preparing Computer-Aided Design (CAD) models and material textures, then constructing a realistic inspection scene incorporating domain-randomized camera settings, lighting, and background elements. The generated data is assessed for effectiveness in both supervised and unsupervised defect detection tasks. Additionally, sim-to-real transferability is examined, demonstrating that models trained on the generated synthetic data can effectively detect and classify defects in real blade images.

Automated defect detection in aero-engine blades is essential for upholding high safety and performance standards in the aerospace industry1,2. However, the absence of large, annotated datasets that cover various types of defects poses a significant challenge in creating and evaluating effective inspection models. This study aims to fill this gap by introducing a high-quality synthetic dataset specifically designed for detecting defects in aero-engine blades. In the industrial domain, data scarcity poses a substantial obstacle to the training and evaluation of AI models3. The infrequency of defective samples, coupled with their bias towards specific defects, leads to data imbalance. Moreover, the process of annotating defects is labor-intensive and demands highly experienced laborers.

Read the full article at Nature.com.