ICD-YOLO:改进YOLOv8的输电线路覆冰检测方法OA
ICD-YOLO:Improved YOLOv8 Method for Detecting Ice Accumulation on Power Transmission Lines
输电线路覆冰作为冬季频发的自然灾害,严重威胁电网安全稳定运行.针对传统巡检方法在复杂地形环境下存在效率低下、实时性差等问题,本文提出一种改进YOLOv8s框架的深度学习模型ICD-YOLO.首先,用EfficientViT混合骨干网络替代传统CNN结构,通过分层注意力机制实现局部特征与全局上下文的动态融合,有效缓解特征稀释问题;其次,设计CEBiFPN多尺度架构,通过双向跨尺度连接和通道-空间注意力机制增强模型对不同厚度覆冰特征的提取能力;最后,提出动态检测头DyHead与Focaler-IoU损失函数的协同优化方案,通过多分支注意力机制实现关键区域的自适应聚焦,降低模型对困难样本的检测难度.实验结果表明,ICD-YOLO与YOLOv8s相 比,mAP@0.5指标从79.8%提升至80.5%,模型参数量从11.1 M降低至6.4 M,每次浮点运算次数从28.7 GFLOPS降低至17.5 GFLOPS.
As a natural disaster frequently occurring in winter,icing on transmission lines seriously threatens the safe and stable operation of power grids.To address the inefficiency and poor real-time performance of traditional inspection methods in complex terrain environments,this paper proposes a deep learning model based on an improved YOLOv8s framework called ICD-YOLO.First,an EfficientViT hybrid backbone network replaces the traditional CNN structure,achieving dynamic fusion of local fea-tures and global context through a hierarchical attention mechanism,effectively alleviating the feature dilution problem.Next,a CEBiFPN multi-scale architecture is designed,enhancing the model's ability to extract icing features of varying thickness through bidirectional cross-scale connections and channel-spatial attention mechanisms.Finally,a collaborative op-timization scheme of the dynamic detection head DyHead and the Focaler-IoU loss function is proposed,enabling adaptive focus on key regions through a multi-branch attention mechanism and reducing the difficulty of detecting challenging samples.Experimental results show that compared with YOLOv8s,the ICD-YOLO improves the mAP@0.5 metric from 79.8%to 80.5%,reduces the model parameters from 11.1 M to 6.4 M,and decreases the number of floating-point operations per calculation from 28.7 GFLOPS to 17.5 GFLOPS.
张旖凡;邓海峰;鲍刚
三峡大学电气与新能源学院,湖北 宜昌 443002三峡大学电气与新能源学院,湖北 宜昌 443002三峡大学电气与新能源学院,湖北 宜昌 443002
信息技术与安全科学
深度学习电力系统输电线路覆冰检测YOLOv8s
deep learningpower systemstransmission line icing detectionYOLOv8s
《计算机与现代化》 2026 (9)
77-85,9
国家自然科学基金资助项目(62476153)
评论