航空学报 > 2023, Vol. 44 Issue (S2): 729283-729283   doi: 10.7527/S1000-6893.2023.29283

基于深度学习的结冰区域和厚度检测方法

苏鑫1,2, 管润程1,2, 王桥3, 苑伟政1,2, 吕湘连1,2, 何洋1,2()   

  1. 1.西北工业大学 空天微纳系统教育部重点实验室,西安 710072
    2.西北工业大学 陕西省微/纳米系统重点实验室,西安 710072
    3.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
  • 收稿日期:2023-07-10 修回日期:2023-07-16 接受日期:2023-08-07 出版日期:2023-08-21 发布日期:2023-08-18
  • 通讯作者: 何洋 E-mail:heyang@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(51875478);航空科学基金(2017ZC53036)

Ice area and thickness detection method based on deep learning

Xin SU1,2, Runcheng GUAN1,2, Qiao WANG3, Weizheng YUAN1,2, Xianglian LYU1,2, Yang HE1,2()   

  1. 1.Key Lab of Micro/Nano Systems for Aerospace,Ministry of Education,Northwestern Polytechnical University,Xi’an 710072,China
    2.Shaan’xi Key Lab of MEMS/NEMS,Northwestern Polytechnical University,Xi’an 710072,China
    3.Key Laboratory of Icing and Anti/De-icing,China Aerodynamics Research and Development Center,Mianyang 621000,China
  • Received:2023-07-10 Revised:2023-07-16 Accepted:2023-08-07 Online:2023-08-21 Published:2023-08-18
  • Contact: Yang HE E-mail:heyang@nwpu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(51875478);Aeronautical Science Foundation of China(2017ZC53036)

摘要:

飞机结冰问题是飞机安全飞行的最大威胁之一,在世界军、民飞机事故中占一半以上,因此对飞机飞行时期结冰状态进行检测尤为重要。目前传统的结冰传感器只能定点检测结冰信息,无法反馈整体区域的结冰状态。提出一种大面积结冰检测方法,利用深度学习算法,建立翼型实时结冰检测模型,以相机采集到的机身图像作为输入,实时检测并分割出翼型结冰区域,计算得到翼型前缘结冰厚度。在自建的结冰数据集中进行验证试验,结果表明:该算法的平均准确率可达99.32%,平均交并比可达93.20%,并且在1.5 m的距离下理论空间分辨率约为0.4 mm,实际测量误差在1 mm以内,实时性检测帧率为4 frame/s。试验证明该方法具有较好的检测效果,有望为飞机结冰检测和地面冰风洞试验奠定研究基础。

关键词: 深度学习, 语义分割, 结冰检测, 厚度检测, UNet

Abstract:

Aircraft icing has become one of the biggest threats to safe flight of aircraft, accounting for more than half of military and civil aircraft accidents in the world. Therefore, it is extremely important to detect the icing conditions during flight. Traditional icing sensors can only detect icing information at fixed points and cannot provide feedback on the icing status of overall area. In this paper, a large-area icing detection method is proposed, and a real-time airfoil icing detection model is established by using deep learning algorithm. Airfoil icing area is detected and segmented in real time using the fuselage image captured by cameras as input, and the icing thickness of the airfoil leading edge is calculated. The experimental results from the self-built icing dataset show that the average accuracy of the algorithm is 99.32%, the mean intersection over union is 93.20%, the theoretical spatial resolution is approximately 0.4 mm at a distance of 1.5 m, the actual measurement error is within 1 mm, and the real-time detection frame rate is 4 frame/s. The experiments demonstrate that this method has excellent detection performance, and potentially lay research foundation for aircraft icing detection and ground icing wind tunnel tests.

Key words: deep learning, semantic segmentation, ice detection, thickness detection, UNet

中图分类号: