摘要
针对电力物联网中设备表面缺陷自动化检测难题(如破损、污损及人为违规操作导致的缺陷),提出一种面向边缘计算设备的轻量级SSD检测算法。该算法通过3个关键技术创新实现高效检测。首先,在MobileNetV2的瓶颈结构中引入密集连接机制,动态增强图像特征表达能力;其次,基于Non-Local注意力机制构建跨层注意力隐式特征金字塔网络(CL-IFPN),通过与MobileNetV2-SSD的深度融合显著提升小缺陷检测能力;最后,通过在卷积层添加特征融合模块并采用QFL函数,强化不同尺度缺陷的预测精度及正负样本训练平衡性。实验结果表明:在公共数据集VOC2007上,所提算法以79.62%的mAP检测精度和36帧/s的检测速度表现优于同类算法;在自建电力器件缺陷数据集上,检测性能进一步提升至95.19%的检测精度和24帧/s的检测速度,充分验证了算法在电力设备缺陷检测场景的实用价值。所提算法为边缘计算环境下的电力物联网设备智能运维提供了有效的技术解决方案。
To address the challenge of automated surface defect detection(e.g.,damage,stains,and defects from human violations)in the power IoT,a lightweight SSD detection algorithm for edge computing devices was proposed.The proposed algorithm aimed to achieved efficient detection through three key innovations.Firstly,a dense connection mechanism was introduced into the bottleneck structure of MobileNetV2 to enhance image feature representation dynamically.Secondly,a cross layer attention mechanism implicit feature pyramid network(CL-IFPN)based on No-Local attention mechanism was constructed,and its deep integration with MobileNetV2-SSD significantly improved small-defect detection.Finally,a feature fusion module was added to the convolutional layer,and the QFL function was used to boost prediction accuracy of defects at different sizes and the balance of positive and negative sample training.Experimental results showed that on the public dataset VOC2007,the proposed algorithm achieved a detection accuracy of 79.62%and a speed of 36 frames per second,outperforming similar algorithms.On the self-built power device defect dataset,the detection accuracy reached 95.19%and a speed of 24 frames per second,demonstrating the algorithm′s practicality in power device defect detection.The proposed algorithm offered an effective technical solution for intelligent operation and maintenance of power IoT devices in edge computing environments.
作者
蔡宇翔
陈丽娟
安琪
CAI Yuxiang;CHEN Lijuan;AN Qi(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;Information and Communication Branch,State Grid Fujian Electric Power Co.,Ltd.,Fuzhou 350001,China;Beijing Fibrlink Communications Co.,Ltd.,Beijing 100070,China;Beijing Chuang’an Hengyu Technology Co.,Ltd.,Beijing 100070,China)
出处
《郑州大学学报(工学版)》
北大核心
2026年第1期140-146,共7页
Journal of Zhengzhou University(Engineering Science)
基金
福建省自然科学基金资助项目(2024J01207)。
