Fault Detection Method for Arresters Based on Lightweight U-Net and Infrared Images
摘要
针对复杂环境下,红外图像存在低对比度、低信噪比、背景干扰较多导致避雷器本体分割不准的问题,本研究提出基于轻量化U-Net的改进语义分割算法.该算法采用倒残差结构(In-verted Residual)与挤压和激励网络(Squeeze-and-Excitation Networks,SENet)融合的方式建立主干网络,并在跳跃连接部分引入高效通道注意力(Efficient Channel Attention,ECA)机制,使网络性能得到增强的同时减少参数量和计算量.与PAN、DeepLabv3+、PSPNet、U-Net等主流模型相比,本研究算法的平均交并比、平均像素准确率、参数量、计算量和推理速度分别为84.72%、99.79%、6.629M、17.028G和50.3f/s,在分割准确率、模型轻量化和推理速度方面综合表现最优.同时,提出基于红外图像底层二进制数据的避雷器本体温度识别方法,避免了传统灰度-温度曲线拟合法带来的拟合误差.最后,基于避雷器表面温度分布曲线进行故障诊断,并通过加压试验验证了其有效性.该方法可为避雷器无接触、带电故障检测提供重要技术支持,促进电网智能化巡检的发展.
In response to the challenges posed by complex environments where infrared images suffer from low contrast,low signal-to-noise ratio,and significant background interference leading to inaccurate segmentation of arresters,the author proposes an improved semantic segmentation algorithm based on a lightweight U-Net.The algorithm integrates an inverted residual structure with SENet to establish the backbone network.ECA are introduced in the skip-connection part to enhance network performance while reducing parameters and computational load.Compared to mainstream models such as PAN,Deep-Labv3+,PSPNet,and U-Net,the proposed algorithm achieves superior performance in terms of MI-oU,MPA,parameters,computational load,and inference speed,with values of 84.72%,99.79%,6.629 M,17.028 G,and 50.3 frames per second respectively.Additionally,an arrester temperature recognition method based on infrared image binary data is proposed to avoid fitting errors associated with traditional grayscale-temperature curve fitting methods.Finally,fault diagnosis based on arrester surface temperature distribution curves is conducted and experimentally validated for its effectiveness.This ap-proach provides crucial technical support for non-contact,live fault detection of arresters,thereby pro-moting the development of intelligent grid inspection.