Insulator self-explosion detection in transmission line based on CenterNet fusing lightweight features
-
摘要:
输电线路智能化巡检是新一代电力系统建设的必然要求。当前，基于深度学习的检测模型由于参数量过大，使得利用无人机（UAV）进行边缘部署较困难。为使无人机可搭载轻量级模型实现输电线路中具有自爆缺陷绝缘子的识别，提出了一种轻量级CenterNet-GhostNet的目标检测网络。对模型主干特征提取网络进行轻量化处理，利用计算成本较低的GhostNet提取自爆缺陷绝缘子的多层次特征，降低模型复杂度；引入增强感受野模块(RFB)增强特征表达能力，提升模型对小目标特征信息的注意力；构建特征融合模块，将低层特征信息和高层特征信息有效融合以输出更完整的特征图，提高缺陷识别精度。利用迁移学习参数共享，结合冻结与解冻训练相结合的模型训练策略，缓解网络因小样本数据集而产生的泛化能力不足问题。基于构建的输电线路自爆缺陷绝缘子数据集对所提方法进行验证，实验结果表明：相比原始CenterNet，所提方法的AP50、AP75和AP50:95分别提升至0.86、0.74和0.63，模型参数量由124.61 ×106减少至64.2 ×106，可实现复杂环境下的自爆缺陷绝缘子检测，提高了基于无人机的输电线路巡检精度与速度。
Abstract:Intelligent inspection of transmission lines is an inevitable requirement for the construction of a new generation of power systems. At present, the detection model based on deep learning has too many parameters, which makes it difficult to deploy unmanned aerial vehicles (UAVs) at the edge. In order to enable the UAV to carry a lightweight model to identify insulators with self-explosion defects in transmission lines, a lightweight CenterNet-GhostNet target detection network was proposed. Firstly, the backbone feature extraction network of the model received lightweight treatment, and the multi-level features of insulators with self-explosion defects were extracted by using GhostNet with low computational costs, so as to reduce the complexity of the model. Then, the enhanced receptive field block (RFB) was introduced to enhance the ability of feature expression and enhance the attention of the model to the feature information of small targets. Finally, a feature fusion module was constructed to effectively fuse the low-level feature information and high-level feature information, so as to output a more complete feature map and improve the accuracy of defect recognition. The model training strategy of sharing transfer learning parameters and combining freezing and thawing training was used, so as to avoid insufficient generalization ability of the network caused by a small sample dataset. Based on the constructed dataset of insulators with self-explosion defects in transmission lines, the proposed method was verified. The experimental results show that compared with the original CenterNet, AP50, AP75, and AP50:95 of the proposed method are increased to 0.86, 0.74, and 0.63, respectively, and the number of model parameters is reduced from 124.61 ×106 to 64.2 ×106. Therefore, the proposed method can detect insulators with self-explosion defects in complex environments and improve the inspection accuracy and speed of transmission lines based on UAVs.
-
图 2 GhostNet-RFB轻量化特征提取网络
Figure 2. Lightweight feature extraction network of GhostNet-RFB
图 7 改进后轻量级CenterNet目标检测网络结构
Figure 7. Structure of improved lightweight CenterNet target detection network
图 8 轻量级CenterNet-GhostNet训练损失函数曲线
Figure 8. Training loss function curves of lightweight CenterNet-GhostNet
图 9 使用迁移学习和未使用迁移学习模型训练损失曲线
Figure 9. Training loss curves with and without transfer learning model
图 10 改进前后CenterNet可视化检测结果
Figure 10. Visual detection results of CenterNet before and after improvement
表 1 GhostNet特征层输出尺度与网络参数结构
Table 1. Output scale and network parameter structure of GhostNet feature layer
输入尺度/(像素×像素×像素) 操作 步长 512×512×3 Conv2d 3×3 2 256×256×16 G-bneck(×2) 1,2 128×128×24 G-bneck(×2) 1,2 64×64×40 G-bneck(×2) 1,2 32×32×80 G-bneck(×4) 1,1,1,1 32×32×112 G-bneck(×2) 1,2 16×16×160 G-bneck(×4) 1,1,1,1 16×16×160 Conv2d 3×3 1 
下载: 导出CSV
表 2 主干网络对比指标
Table 2. Comparison indexes of backbone network
主干网络 参数量 模型大小 SqueezeNet 6.9×106 28.3×106 MobileNetV3 5.8×106 26.8×106 ShuffleNetV2 5.3×106 23.5×106 GhostNet 5.18×106 19.77×106
下载: 导出CSV
表 3 消融实验评价指标
Table 3. Evaluation indexes of ablation experiment
特征提取 特征融合 AP50 AP75 AP50:95 参数量 FLOP GhostNet RFB 0.68 0.62 0.59 124.61×106 19.1×106 √ 0.71 0.64 0.62 47.5×106 15.7×106 √ 0.76 0.71 0.64 48.2×106 15.9×106 √ 0.81 0.69 0.67 52.4×106 17.3×106 √ √ √ 0.86 0.74 0.63 64.2×106 18.8×106
下载: 导出CSV
表 4 不同检测方法的对比实验评价指标
Table 4. Comparative experimental evaluation indexes of different detection algorithms
方法 骨干网络 AP50 AP75 帧率/(帧·s−1) Faster R-CNN VGG16 0.77 0.62 16.3 Mask R-CNN ResNet50 0.70 0.67 13 YOLOX Darknet53 0.76 0.75 32.3 YOLOv4-Tiny CSPDarknet53 0.73 0.68 28.4 MobileNet-SSD MobileNet 0.82 0.71 30.6 本文方法 GhostNet 0.86 0.74 38.2
下载: 导出CSV
-
[1] 赵振兵. 电力视觉技术[M]. 北京: 中国电力出版社, 2020: 84-90.ZHAO Z B. Computer vision technology in electric power system[M]. Beijing: China Electric Power Press, 2020: 84-90(in Chinese). [2] 赵振兵, 张薇, 翟永杰, 等. 电力视觉技术的概念、研究现状与展望[J]. 电力科学与工程, 2020, 36(1): 1-8. doi: 10.3969/j.ISSN.1672-0792.2020.01.001ZHAO Z B, ZHANG W, ZHAI Y J, et al. Concept, research status and prospect of electric power vision technology[J]. Electric Power Science and Engineering, 2020, 36(1): 1-8(in Chinese). doi: 10.3969/j.ISSN.1672-0792.2020.01.001 [3] 陶显, 侯伟, 徐德. 基于深度学习的表面缺陷检测方法综述[J]. 自动化学报, 2021, 47(5): 1017-1034.TAO X, HOU W, XU D. A survey of surface defect detection methods based on deep learning[J]. Acta Automatica Sinica, 2021, 47(5): 1017-1034(in Chinese). [4] 刘洋, 陆倚鹏, 高嵩, 等. 边缘检测在盘形悬式瓷绝缘子串红外图像上的应用[J]. 电瓷避雷器, 2020(1): 198-203.LIU Y, LU Y P, GAO S, et al. Edge detection on infrared image of high voltage porcelain disc type suspension insulator strings[J]. Insulators and Surge Arresters, 2020(1): 198-203(in Chinese). [5] 赵振兵, 徐磊, 戚银城, 等. 基于Hough检测和C-V模型的航拍绝缘子自动协同分割方法[J]. 仪器仪表学报, 2016, 37(2): 395-403. doi: 10.3969/j.issn.0254-3087.2016.02.021ZHAO Z B, XU L, QI Y C, et al. Automatic co-segmentation method for aerial insulator based on Hough detection and C-V model[J]. Chinese Journal of Scientific Instrument, 2016, 37(2): 395-403(in Chinese). doi: 10.3969/j.issn.0254-3087.2016.02.021 [6] 姚晓通, 刘力, 李致远. 基于Canny边缘特征点的接触网绝缘子识别方法[J]. 电瓷避雷器, 2020(1): 142-148.YAO X T, LIU L, LI Z Y. Identification method of catenary insulator based on Canny edge feature point[J]. Insulators and Surge Arresters, 2020(1): 142-148(in Chinese). [7] 崔昊杨, 张雨阁, 张驯, 等. 基于边端轻量级网络的电力仪表设备检测方法[J]. 电网技术, 2022, 46(3): 1186-1193.CUI H Y, ZHANG Y G, ZHANG X, et al. Detection of power instruments equipment based on edge lightweight network[J]. Power System Technology, 2022, 46(3): 1186-1193(in Chinese). [8] 顾超越, 李喆, 史晋涛, 等. 基于改进Faster-RCNN的无人机巡检架空线路销钉缺陷检测[J]. 高电压技术, 2020, 46(9): 3089-3096.GU C Y, LI Z, SHI J T, et al. Detection for pin defects of overhead lines by UAV patrol image based on improved Faster-RCNN[J]. High Voltage Engineering, 2020, 46(9): 3089-3096(in Chinese). [9] 李鑫, 刘帅男, 杨桢, 等. 基于改进Cascade R-CNN的输电线路多目标检测[J]. 电子测量与仪器学报, 2021, 35(10): 24-32.LI X, LIU S N, YANG Z, et al. Multi-target detection of transmission lines based on improved Cascade R-CNN[J]. Journal of Electronic Measurement and Instrumentation, 2021, 35(10): 24-32(in Chinese). [10] 李瑞生, 张彦龙, 翟登辉, 等. 基于改进SSD的输电线路销钉缺陷检测[J]. 高电压技术, 2021, 47(11): 3795-3802.LI R S, ZHANG Y L, ZHAI D H, et al. Pin defect detection of transmission line based on improved SSD[J]. High Voltage Engineering, 2021, 47(11): 3795-3802(in Chinese). [11] 郝帅, 杨磊, 马旭, 等. 基于注意力机制与跨尺度特征融合的YOLOv5输电线路故障检测[J]. 中国电机工程学报, 2023, 43(6): 2319-2331.HAO S, YANG L, MA X, et al. YOLOv5 transmission line fault detection based on attention mechanism and cross-scale feature fusion[J]. Proceedings of the CSEE, 2023, 43(6): 2319-2331. [12] 许延雷, 梁继然, 董国军, 等. 基于改进CenterNet的航拍图像目标检测算法[J]. 激光与光电子学进展, 2021, 58(20): 2010013.XU Y L, LIANG J R, DONG G J, et al. Aerial image target detection algorithm based on improved CenterNet[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010013(in Chinese). [13] 唐玮, 赵保军, 龙腾. 基于轻量化网络的光学遥感图像飞机目标检测[J]. 信号处理, 2019, 35(5): 768-774.TANG W, ZHAO B J, LONG T. Aircraft detection in remote sensing image based on lightweight network[J]. Journal of Signal Processing, 2019, 35(5): 768-774(in Chinese). [14] 符惠桐, 王鹏, 李晓艳, 等. 面向移动目标识别的轻量化网络模型[J]. 西安交通大学学报, 2021, 55(7): 124-131. doi: 10.7652/xjtuxb202107014FU H T, WANG P, LI X Y, et al. Lightweight network model for moving object recognition[J]. Journal of Xi’ an Jiaotong University, 2021, 55(7): 124-131(in Chinese). doi: 10.7652/xjtuxb202107014 [15] 汝承印, 张仕海, 张子淼, 等. 基于轻量级MobileNet-SSD和MobileNetV2-DeeplabV3+ 的绝缘子故障识别方法[J]. 高电压技术, 2022, 48(9): 3670-3679.RU C Y, ZHANG S H, ZHANG Z M, et al. Fault identification method for high voltage power grid insulator based on lightweight MobileNet-SSD and MobileNetV2-DeeplabV3+ network[J]. High Voltage Engineering, 2022, 48(9): 3670-3679(in Chinese). [16] 王盛洋. 基于迁移学习的轻量化网络目标检测算法研究[D]. 西安: 西安电子科技大学, 2021: 1-15.WANG S Y. Research on lightweight network object detection algorithm based on transfer learning[D]. Xi’an: Xidian University, 2021: 1-15(in Chinese). [17] HAN K, WANG Y H, TIAN Q, et al, GhostNet: More features from cheap operations[EB/OL].(2020-03-13)[2022-07-01] [18] 奉志强, 谢志军, 包正伟, 等. 基于改进YOLOv5的无人机实时密集小目标检测算法[J]. 航空学报, 2023, 44(7): 327106.FENG Z Q, XIE Z J, BAO Z W, et al. Real-time dense small object detection algorithm for UAV based on improved YOLOv5[J]. Acta Aeronautica et Astronautica Sinica, 2023, 44(7): 327106(in Chinese). [19] 李晨瑄, 顾佼佼, 王磊, 等. 多尺度特征融合的Anchor-Free轻量化舰船要害部位检测算法[J]. 北京航空航天大学学报, 2022, 48(10): 2006-2019.LI C X, GU J J, WANG L, et al. Warship’s vital parts detection algorithm based on lightweight Anchor-Free network with multi-scale feature fusion[J]. Journal of Beijing University of Aeronautics and Astronautics, 2022, 48(10): 2006-2019(in Chinese) . [20] 马富齐, 王波, 董旭柱, 等. 面向输电线路覆冰厚度辨识的多感受野视觉边缘智能识别方法研究[J]. 电网技术, 2021, 45(6): 2161-2169.MA F Q, WANG B, DONG X Z, et al. Receptive field vision edge intelligent recognition for ice thickness identification of transmission line[J]. Power System Technology, 2021, 45(6): 2161-2169(in Chinese). [21] 史彩娟, 陈厚儒, 葛录录, 等. 注意力残差多尺度特征增强的显著性实例分割[J]. 图学学报, 2021, 42(6): 883-890.SHI C J, CHEN H R, GE L L, et al. Salient instance segmentation via attention residual multi-scale feature enhancement[J]. Journal of Graphics, 2021, 42(6): 883-890(in Chinese). [22] 宋立业, 刘帅, 王凯, 等. 基于改进EfficientDet的电网元件及缺陷识别方法[J]. 电工技术学报, 2022, 37(9): 2241-2251.SONG L Y, LIU S, WANG K, et al. Identification method of power grid components and defects based on improved EfficientDet[J]. Transactions of China Electrotechnical Society, 2022, 37(9): 2241-2251(in Chinese). -
下载:
点击查看大图
计量
- 文章访问数: 280
- HTML全文浏览量: 87
- PDF下载量: 13
- 被引次数: 0
