基于动态特征融合的机翼结冰区域检测算法  

Wing Icing Area Detection Algorithm Based on Dynamic Feature Fusion

作　　者：张德银[1] 李俊佟 张裕尧 吴章辉 何骏翔 ZHANG Deyin;LI Juntong;ZHANG Yuyao;WU Zhanghui;HE Junxiang(Institute of Electronic and Electrical Engineering,Civil Aviation Flight University of China,Guanghan,Sichuan 618307,China)

机构地区：[1]中国民用航空飞行学院航空电子电气学院,四川广汉618307

出　　处：《自动化应用》2025年第6期40-43,48,共5页Automation Application

基　　金：深圳市科技研发资金项目([2022]47号)。

摘　　要：严重的机翼结冰会对飞行安全构成巨大威胁,结冰越厚、结冰区域越大,对飞行安全的危害程度也越大。目前,机翼结冰区域的检测通常依赖于飞行员目测和单点传感器检测相结合的方法,但这种方法存在结冰边界模糊、难以量化判断以及检测区域范围小的缺陷。为此,提出了一种基于动态特征融合的机翼结冰大范围区域检测算法。在UNeXt神经网络的基础上,引入了残差结构卷积块和动态特征融合模块,构建了一种新的D型结构DR-UNeXt神经网络算法,用于机翼结冰区域的检测。在自建的机翼结冰数据集上对DR-UNeXt新型网络算法进行了实验验证。结果表明,与UNeXt算法相比,DR-UNeXt新型神经网络算法在参数量仅增加12.9%、计算量仅增加31.22 GFLOPs的情况下,IoU从0.941提升至0.957,Dice系数从0.965提升至0.978。引入动态特征融合模块和残差结构卷积块能够更精准地提取机翼结冰区域的边缘信息,提高检测准确率,量化结冰区域,为防冰除冰提供数据支撑。Severe wing icing poses a significant threat to flight safety,and the thicker the icing and the larger the icing area,the greater the hazard to flight safety.Currently,the detection of wing icing areas is typically performed by relying on a combination of pilot visual observation and single-point sensor detection,but this method has drawbacks such as blurred icing boundaries that are difficult to quantify and a limited detection area.To address this issue,a wide-area wing icing detection algorithm based on dynamic feature fusion is proposed.Based on the UNeXt neural network,residual convolutional blocks and dynamic feature fusion modules were introduced to construct a new D-shaped structured neural network algorithm called DRUNeXt,which is used for wing icing area detection.Experimental validation of the new DR-UNeXt network algorithm was conducted on a self-built wing icing dataset.The results show that,compared with the UNeXt algorithm,the IOU was increased from 0.941 to 0.957 and the Dice coefficient from 0.965 to 0.978 by the DR-UNeXt neural network algorithm,with only a 12.9%increase in parameter count and an additional computational cost of just 31.22 GFLOPs.By introducing dynamic feature fusion modules and residual convolutional blocks,more precise edge information of wing icing areas can be extracted,detection accuracy is improved,icing areas are quantified,and data support is provided for anti-icing and de-icing.

关 键 词：D型结构 DR-UNeXt 动态特征融合 结冰区域检测 机翼结冰 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构] V219[自动化与计算机技术—计算机科学与技术]