Microscopic Defect Detection on Aircraft Engine Blades via Improved YOLOv8

Abstract

To fully bring into play the functions of aircraft engine blades, it is indispensable to perform regular inspections of engine blades, which currently rely on inefficient manual visual assessments. While artificial intelligence technology can be utilized, benchmark datasets are not available yet. To tackle these issues, in this work, we first construct two datasets that are collected from real blade defect images at different microscopic magnifications under an electron microscope and a metallographic microscope. Subsequently, we propose an efficient lightweight YOLOv8 framework, incorporating a hierarchical feature fusion module MS-Block for better multi-scale integration, as well as an Efficient Multi-Scale Attention (EMA) and Dilation-wise Residual (DWR) modules to enhance the detection of small targets and replace the loss function with Inner-IoU. The improved YOLOv8 demonstrates a noteworthy increase in mean average precision (mAP), achieving an enhancement of 1.5% on the Electron Microscope Taken (EMT) dataset and 1.8% on the Metallographic Microscope Taken (MMT) dataset compared to the original model. Our approach significantly surpasses the performance of contemporary target detection algorithms, thereby offering a robust solution for microscopic defect detection in aeroengines. This advancement not only streamlines the inspection process but also contributes to the overall safety and reliability of aircraft operations.

Keywords

• Microscopic defect detection
• YOLOv8
• deep learning
• aeroengine