摘要
套管油位异常是导致套管故障、变压器跳闸甚至爆燃的主要原因之一,针对现有超特高压变压器套管油位深度学习检测算法存在的模型结构单一、泛化能力较弱、检测精度无法满足实际运检需求,且难以有效检测、辨识变压器套管油位缺陷等问题,构建了一种改进YOLOv8n的套管油位视觉检测算法:在特征提取方面,基于改进神经卷积网络中的Res2Net模块,设计了一种多尺度空间特征提取模块MSF-RN(multi-scale spatial feature of res2Net),替换YOLOv8n中的原有C2F结构,提升套管油位检测多尺度数据的颗粒度与准确度;在特征金字塔方面,基于Gabor滤波器卷积模块,设计了一种多向空间特征提取模块MDF-GA(multi-direction spatial feature of gabor),有效降低复杂情况下背景信息干扰问题;在检测头的设计方面,提出轻量化自适配归一的卷积检测头SNDH(lightweight switchable normalization convolutional detection head),通过掩蔽卷积来减少各类特征信息之间的差异,提高模型精准度并减少计算量。经在专门针对套管油位的自建数据集上进行实验,结果表明MS-MD-SNDH模型比基线模型精度提升了0.7,参数量及计算量分别下降了42%及56%,可满足实际运检场景的轻量化与高精度需求。
Abnormal oil level in the bushing is one of the main causes of bushing faults,transformer tripping,and even explosions.In response to the limitations of the existing deep learning detection algorithm for bushing oil level in ultra-high voltage transformers,including monolithic model structures,weak generalization capabilities,insufficient detection accuracy for practical maintenance needs,and difficulties in effectively identifying bushing oil level defects,this paper constructs an improved YOLOv8n bushing oil level visual detection algorithm.In terms of feature extraction was designed based on the Res2Net module in improved neural convolutional networks.a multi-scale spatial feature extraction module MSF was designed based on the Res2Net module in improved neural convolutional networks.RN based on the Res2Net module in an improved neural convolutional network,replacing the C2F structure in YOLOv8n and improving the granularity and accuracy of multi-scale data for casing oil level detection.In terms of feature pyramid,a multi-directional spatial feature extraction module MDF-GA(Multi-Directional Spatial Feature of Gabor)was designed based on the Gabor filter convolution module,effectively mitigating background interference issues in complex scenarios.In terms of the design of the detection head,we proposes a new lightweight adaptive normalization convolutional detection head SNDH,which applies masking convolution to reduce the differences between various feature information,improve model accuracy,and reduce computational complexity.Experiments were conducted on a self built dataset specifically designed for casing oil levels,and the results showed that the MS-MD-SNDH model improved accuracy by O.7 compared to the baseline model,while reducing parameter and computational complexity by 42%and 56%,respectively.
作者
刘畅
王乐
胡嘉祥
高淼
崔超
梁贵书
LIU Chang;WANG Le;HU Jiaxiang;GAO Miao;CUI Chao;LIANG Guishu(State Grid Hebei Extra High Voltage Company,Shijiazhuang 050051,China;State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources(NCEPU),Baoding 071000,China)
出处
《河北电力技术》
2025年第5期25-31,共7页
Hebei Electric Power
基金
国家自然基金(51177048)。
