| 引用本文: | 韩畅,韩笑,陈虹,钟杰,戈洋,曹灿,马杰.规范透视场景的半监督目标检测及其在保护压板巡检上的应用[J].电力自动化设备,2023,43(7): |
| HAN Chang,HAN Xiao,CHEN Hong,ZHONG Jie,GE Yang,CAO Can,MA Jie.Semi-supervised object detection for normative perspective scenes and its application on protective connecting pieces[J].Electric Power Automation Equipment,2023,43(7): |
|
| 摘要: |
| 如何快速准确地对继电保护压板的异常状态进行识别，是变电站二次设备巡检工作中亟待解决的技术难题。基于深度学习的通用目标检测算法在向诸如继电保护屏压板检测等特殊化场景的迁移中，不能够充分利用保护屏场景中的规范透视先验特征；此外，人工标注大数据集的困难性一直以来都是通用检测模型迁移至特殊场景时的挑战。针对上述问题，提出了一种适用于保护压板规范化分布特征的半监督目标检测模型，该模型根据压板识别场景的特点对模型框架进行了一系列适应性改进。在模型的半监督训练阶段，使用一致性正则化方法生成伪标签，并基于保护屏压板图像特征，以边缘吸附和点阵行列拟合等方式，优化或剔除伪标签，从而突破了数据标注困难性带来的限制。改进后的模型达到平均精度均值为98.12%的应用级精度，并额外输出图像的逆透视变换参量。该模型被应用于便携式智能终端，辅助工作人员进行继电保护压板状态的巡检工作；模型输出的逆透视变换参量，也可为3D人机交互等下游视觉任务提供技术支撑。 |
| 关键词: 继电保护压板 电力系统智慧化 半监督学习 目标检测算法 逆透视变换 |
| DOI：10.16081/j.epae.202209026 |
| 分类号:TP391.41 |
| 基金项目:江苏省自然科学基金资助项目(BK20181021) |
|
| Semi-supervised object detection for normative perspective scenes and its application on protective connecting pieces |
|
HAN Chang1, HAN Xiao2, CHEN Hong1, ZHONG Jie3, GE Yang1, CAO Can3, MA Jie3
|
|
1.School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China;2.School of Electrical Engineering, Nanjing Institute of Technology, Nanjing 211167, China;3.Lianyungang Power Supply Company of State Grid Jiangsu Power Co.,Ltd.,Lianyungang 222000, China
|
| Abstract: |
| Efficient and accurate identification of abnormal status of protective connecting piece is a pressing technical challenge in the inspection of substation secondary equipment. Deep learning-based generic target detection algorithms are not able to take full advantage of the canonical perspective priori features in protection screen scenarios when migrating to specialization scenarios such as relay protection panel connecting piece detection. The difficulty of manually labeling large data sets has been a challenge when migrating generic detection models to special scenarios. To address the above issues, a semi-supervised object detection model adapted to the normative distribution features is proposed, and a series of adaptations have been made to the model according to the connecting pieces identification scenes. In the semi-supervised training phase of the model, pseudo-labels are generated by using the consistent regularization method, and optimized or rejected by using edge adsorption and dotted ranks fitting based on features of protective screen, thus breaking through the limitations imposed by the difficulty of data annotation. The improved model can achieve an application-level accuracy of 98.12% mean average precision and can additionally output the inverse perspective transformation parameters of the image. The model is applied to a portable intelligent terminal to assist staff in connecting piece inspection. The inverse perspective transformation parameters output by the model can also provide technical support for downstream vision tasks such as 3D human-computer interaction. |
| Key words: protective connecting pieces intellectualization of power systems semi-supervised learning object detection algorithm inverse perspective transformation |
