复杂环境下玻璃绝缘子自爆的实时检测方法  

Real--Time Detection Method for Self-Detonation of Glass Insulator in Complex Environment

作　　者：冯世凯 陈千 FENG Shikai;CHEN Qian(School of Management,Hefei University of Technology,Hefei 230009,China;Key Laboratory of Process Optimization&Intelligent Decision-Making,Ministry of Education,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学管理学院,合肥230009 [2]合肥工业大学过程优化与智能决策教育部重点实验室,合肥230009

出　　处：《电瓷避雷器》2024年第5期135-144,共10页Insulators and Surge Arresters

基　　金：国家自然科学基金(编号:91546108);安徽省科技重大专项(编号:201903a05020020)。

摘　　要：针对无人机巡检电力玻璃绝缘子过程中,无人机的计算能力和存储资源有限,且在黑暗、光照、噪声干扰、特殊天气等复杂环境下拍摄的玻璃绝缘子自爆缺陷图片难以检测的问题。提出一种适用于复杂环境下玻璃绝缘子自爆的实时检测方法。该方法以具有轻量化特点的YOLOX-s目标检测模型为基础。首先,采用多阶段迁移学习策略提升模型整体性能。其次,通过新增一层浅层特征进行特征融合,使得融合后特征的位置信息更加丰富,更利于自爆缺陷小目标的定位。同时利用Focal Loss专注困难样本学习,提高模型在复杂环境下的鲁棒性。最后,采用SIOU计算定位损失以提升模型推理精度。在包含多种复杂环境的玻璃绝缘子自爆数据集上进行实验。实验证明,该方法可以较好的实现复杂环境下玻璃绝缘子及其自爆缺陷的实时检测,mAP0.85高达96.7%,相比于原始YOLOX-s模型提升约6%,而且简单、轻量,适合部署于无人机进行实时检测。Aiming at the problem that the UAV's computing power and storage resources are limited during the inspection of power glass insulators,and the pictures of self-explosion defects of glass insulators taken in complex environments such as darkness,illumination,noise interference and special weather are difficult to detect.A real-time target detection method suitable for self-explosion defect detection of glass insulators in complex environment is proposed.The method is based on YOLOX-s target detection model with lightweight characteristics.Firstly,the multi-stage transfer learning strategy is adopted to improve the overall performance of the model.Secondly,by adding a layer of shallow features for feature fusion,the location information of the fused features is more abundant,which is more conducive to the positioning of small targets of self-explosion defects.Then,Focal Loss is used to focus on learning difficult samples and improve the robustness of the model in complex environment.Finally,SIOU is used to calculate the positioning loss of the model to improve the reasoning accuracy of the model.Experiments are carried out on the self-explosion data set of glass insulators with various complex backgrounds.Experiments show that the method can realize the real-time detection of glass insulators and their self-explosion defects in complex environments,with mAP0.85 as high as 96.7%,which is about 6%higher than that of the original YOLOX-s model.It is simple and lightweight which is suitable for real-time detection in unmanned aerial vehicles.

关 键 词：玻璃绝缘子 自爆 实时检测 复杂环境 

分 类 号：TM216.4[一般工业技术—材料科学与工程]