Abstract: Remainder particles in closed electronic equipment equipped in spacecraft bring huge hidden danger to the flight safety of spacecraft.Since remainder particles are in small size,and even the morphological structure of the remainder particles is highly similar to the general components in equipment,and remainder particles are easily covered by other components,the current methods used to detect remainder particles can cause false detection and missed detection frequently.To resolve these problems,a Remainder Particle Detection Network (RPDN) was proposed to detect remainder particles in closed electronic equipment based on convolution-inverted residual and combined attention mechanism.A convolution-inverted residual module was built to ensure the integrity of the remainder particles' fine-grained feature.Then,the combined attention mechanism was proposed to enhance the representativeness of remainder particles feature.The objects were predicted from multiple dimensions by combining multi-scale feature fusion module and object detection layer.The experimental results showed that RPDN had achieved good effect in all evaluation indicators,the mAP of the proposed method reached to 92.16%,and the detection efficiency reached 13FPS.It realized efficient and accurate detection of remainder particles in closed electronic equipment equipped in spacecraft.

Key words: spacecraft, closed electronic equipment, remainder particles detection, convolution-inverted residual module, combined attention mechanism

摘要： 航天器密闭电子设备内腔多余物给航天器飞行安全带来了巨大隐患。由于多余物体积小、与设备内常规组件形态结构相似且易被其他组件遮挡,采用现有的方法对其进行检测时误检、漏检频发。为解决上述问题,提出一种基于卷积—反残差和组合注意力机制的航天器密闭电子设备多余物检测网络RPDN。首先,网络通过构建卷积—反残差模块,保证了多余物细粒度特征的完整性;其次,设计组合注意力机制,增强了多余物特征的表征能力;最后,结合多尺度特征融合模块与目标检测层从多维度进行目标预测。实验结果表明RPDN在各项评价指标上均取得了良好的效果,mAP达到92.16%,检测效率达到了13FPS,实现了航天器密闭电子设备内腔多余物高效、精准检测。

关键词: 航天器, 密闭电子设备, 多余物检测, 卷积—反残差模块, 组合注意力机制