摘要
针对目前人工检测航空发动机叶片表面缺陷存在检测效率低、易漏检的问题,提出一种轻量的智能缺陷检测模型YOLOv5-GA。首先,在主干网络中添加了GhostConv模块和C3Ghost模块来减少参数量和计算量,使网络更加轻量。然后,在颈部网络中融合了渐进特征金字塔算法(AFPN)来提高网络对小目标的检测效果。试验结果表明,在航空发动机叶片的缺陷识别上,本文算法不仅mAP达到了92.6%,较基准网络提升了4.6个百分点,而且训练后的模型大小仅为9 MB,较基准网络减小了38%。在NEU-DET数据集上,mAP达到了77%,优于其他网络。训练后的模型大小也明显减小。因此,所提出的网络具有轻量、高效、可靠、好的泛化能力等特点,可以有效检测航空发动机叶片的主要缺陷。
Addressing the challenges of low efficiency and potential oversight in artificial detection of surface defects on aero-engine blades,this paper introduces a novel lightweight intelligent defect detection model,termed YOLOv5-GA.The model incorporates a GhostConv module and C3Ghost into the backbone network to minimize parameters and computational load,thereby enhancing its lightweight nature.Furthermore,the integration of an asymptotic feature pyramid network(AFPN)into the neck network enhances the model’s capability to detect small targets.Experimental findings demonstrate that in the domain of aircraft engine blade defect recognition,the proposed algorithm not only achieves an mAP of 92.6%,a 4.6-percentage-point enhancement over the baseline network but also reduces the model size to a mere 9 MB,reflecting a 38%reduction compared to the baseline.Additionally,on the NEU-DET dataset,the model achieves an mAP of 77%,outperforming other networks while significantly reducing model size.Thus,the proposed network boasts lightweight,efficient,and reliable characteristics,facilitating the effective detection of critical defects in aero-engines.
作者
罗金超
郑波
LUO Jinchao;ZHENG Bo(Civil Aviation Flight University of China,Guanghan 618307,China)
出处
《航空制造技术》
北大核心
2025年第23期50-58,共9页
Aeronautical Manufacturing Technology
基金
中央高校基本科研业务费(24CAFUC04016)。
