基于改进YOLOv8的航空发动机叶片表面缺陷检测  

Detection of Surface Defects on Aircraft Engine Blades Based on Improved YOLOv8

作　　者：李文龙 王欣威[1] 慕丽[1] LI Wenlong;WANG Xinwei;MU Li(School of Mechanical Engineering,Shenyang University of Technology,Shenyang 110159,China)

机构地区：[1]沈阳理工大学机械工程学院,沈阳110159

出　　处：《组合机床与自动化加工技术》2024年第12期46-50,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(51934002)。

摘　　要：针对航空发动机叶片表面缺陷的复杂性,检测效率和精度不高的问题,提出了一种改进的基于注意力机制的YOLOv8s航空发动机叶片表面缺陷检测方法。通过将EIoU替换为CIoU作为算法的损失函数。在提高边界框回归速率和目标定位精度的同时,改善数据集中的质量不平衡问题。在主干特征网络(Backbone)中嵌入EMA注意力模块,以增强对关键特征的提取,提高模型的检测准确性。使用自建的航空发动机叶片数据集对网络进行训练和测试。试验结果表明,YOLOv8s-EMA网络的平均检测精确度达到了98.7%。相较于Faster-RCNN和YOLOv5s等目前主流的目标检测模型,平均检测精确度分别提高了2.1%和3.0%,FPS也有显著提升。证明了该方法在航空发动机叶片表面缺陷检测中具有更高的精度,取得了良好的检测效果。In response to the reflective nature and complex surface defects of aircraft engine blades,this paper proposes an improved aircraft engine blade surface defect detection method based on an attention mechanism and enhanced YOLOv8s.By replacing EIoU with CIoU as the algorithm′s loss function.Simultaneously,improvements were made in enhancing the accuracy of boundary box regression and target localization,while addressing the issue of data quality imbalances.The integration of an EMA attention module within the main feature network(Backbone)aimed to amplify the extraction of crucial features and enhance the model′s detection accuracy.The network was trained and tested using a proprietary dataset of aircraft engine blades.Experimental results demonstrate that the average detection accuracy of the YOLOv8s-EMA network reached 98.7%.Compared to current mainstream target detection models such as Faster-RCNN and YOLOv5s,the average detection accuracy improved by 2.1%and 3.0%respectively,with a significant increase in FPS.This evidence validates the method′s higher precision in detecting surface defects on aircraft engine blades,achieving commendable detection results.

关 键 词：缺陷检测 叶片 YOLOv8s EMA注意力机制 EIoU 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]