基于轻量级网络的飞机蜂窝结构积水缺陷检测

Water Ingress Detection of Aircraft Honeycomb Structure Based on Lightweight Network

摘要:

复合材料蜂窝结构在飞机服役过程中产生的积水缺陷在日常维护和检修过程中依赖人工、检测效率低、自动化程度低，若未能及时发现将严重威胁飞行安全。针对该问题，结合实际检修场景中使用的移动或嵌入式设备算力有限的情况，设计了一种融合通道注意力和倒残差算法的模块SE-IR,进一步搭建了基于SE-IR模块的轻量级网络SE-IR LCNN，尽可能地在保证网络检测准确率的同时减小网络的参数量。为了验证所提轻量级网络的有效性、使用数字X射线摄影设备获取蜂窝结构及其积水缺陷数字化图像并制成数据集。在该数据集上的实验结果表明，所提轻量级网络的分类准确率为99.20%，可有效筛选出飞机蜂窝结构的积水缺陷。相较于经典网络ResNet-50和VGG-16，所提网络的准确率分别提升了9.6%和3.66%、参数量仅为ResNet-50参数量的1/10、VGG-16参数量的1/50。

Abstract:

The water ingress defects of composite honeycomb structure produced in the service process of aircraft will seriously threaten flight safety.The detection of water ingress defects rely on manual work, low detection efficiency and low degree of automation in the daily maintenance and overhaul process. Aiming at this problem, considering the limited computing power of mobile or embedded devices used in actual maintenance scenarios, a module SE-IRthat integrates sequeeze and excitation block and inverted residual algorithm is designed, and a lightweight network SE-IR LCNN based on SE-IR module is further built. As much as possible to ensure the accuracy of network detection while reducing the number of network parameters. In order to verify the effectiveness of the proposed lightweight network, digital X-ray photography equipment is used to obtain digital images of honeycomb structures and their water defects and make data sets. The experimental results on this dataset show that the classification accuracy of the proposed lightweight network is 99.20 %, which can effectively screen out the water accumulation defects of aircraft honeycomb structure. Compared with the classical network ResNet-50 and VGG-16, the accuracy of the proposed network is increased by 9.6 % and 3.66 % respectively, and the number of parameters is only 1 / 10 of the number of parameters of ResNet-50 and 1 / 50 of the number of parameters of VGG-16.