一种改进的CenterNet机翼结冰检测方法  

An Improved CenterNet Method for Wing Icing Detection

作　　者：王一帆 魏家田 左承林 周文俊 熊浩 赵荣 彭博[1] 王杨[1] WANG Yifan;WEI Jiatian;ZUO Chenglin;ZHOU Wenjun;XIONG Hao;ZHAO Rong;PENG Bo;WANG Yang(School of Computer Science,Southwest Petroleum University,Chengdu 610500,P.R.China;Key Laboratory of Icing and Anti/De‑icing of China Aerodynamics Research and Development Center,Mianyang 621000,P.R.China)

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]中国空气动力研究与发展中心结冰与防除冰重点实验室,绵阳621000

出　　处：《Transactions of Nanjing University of Aeronautics and Astronautics》2023年第6期703-713,共11页南京航空航天大学学报（英文版）

基　　金：supported by the Key Laboratory of Icing and Anti/De-icing of China Aerodynamics Research and Development Center (CARDC) (No. IADL20210203);the Natural Science Foundation of Sichuan, China (No.2023NSFSC1393);the Scientific Research Starting Project of Southwest Petroleum University (SWPU) (No.2021QHZ001);the National Natural Science Foundation of China (No.52006235)。

摘　　要：在高空飞行环境下,机翼表面的积冰可能会改变其空气动力学特性并进一步降低升力,因此机翼结冰的检测显得尤为关键。为克服现有积冰检测技术通常依赖于操作人员的经验判断或需要昂贵的工程实施及硬件开发成本的局限,本文提出了一种基于CenterNet模型的旋转目标检测方法 RA-CenterNet。针对风洞实验数据集中呈现的特定积冰区域方向问题,设计了一种新颖的角度预测分支网络,有效实现了对旋转目标的精确校准。此外,为了提升神经网络在识别冰形边界时的特征提取能力,研究中还融合了卷积注意力模块(Convolutional block attention module, CBAM)。通过与其他旋转目标检测方法以及基准网络进行的一系列对比实验验证了RA-CenterNet方法的性能。实验结果表明,RA-CenterNet算法在主流的旋转目标检测算法中显示出明显的竞争优势,证明了其在积冰检测领域的应用潜力。Aircraft wing icing detection is a crucial task during high-altitude flights because ice accumulation on the leading edge of wings can change their aerodynamic shape and reduce lift capacity.This paper proposes a rotated object detection method called RA-CenterNet,based on the CenterNet model,to overcome the limitations of existing icing detection approaches that either rely on operator experience or require high engineering implementation and hardware development costs.To address the specific icing area directions presented in wind tunnel experimental datasets,a novel angle prediction branch network that enables precise calibration of rotated targets is designed.Additionally,the convolutional block attention module(CBAM)is incorporated to enhance the feature extraction ability of the neural network for ice-shaped boundaries.Comparative experiments are conducted to validate the performance of the proposed method against other rotated object detection approaches and the baseline network.The results demonstrate that our RA-CenterNet method has a significant competitive advantage over the mainstream rotation-based object detection algorithms.

关 键 词：机翼结冰 深度学习 旋转目标检测 无锚点 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]