基于深度注意力残差网络的电表破损检测研究
Research on Electric Meter Damage Detection Based on Deep Attention Re-sidual Network
陈彪 1许建明 2李榕桂1
作者信息
- 1. 福建网能科技开发有限责任公司,福建 福州 350003
- 2. 国网信息通信产业集团有限公司福州分公司,福建 福州 350003
- 折叠
摘要
当前,人工检测是拆回电能表的外观破损检测主流方法,其严重制约了拆回电能表的识别效率,为此提出一种基于注意力机制和残差思想的深度卷积神经网络模型RAN,依据注意力机制,在RAN中构建通道和空间注意力模块,以增强网络的特征提取能力,并借助残 差思想避免注意力模块造成的特征值衰减.以火烧表、水浸表、端子异常表、显示屏破损表、外观破裂表、正常表为研究对象.RAN平均给识别率为94.58％,比ResNet101、VGG16、MobileNet和ShuffleNet提升了0.32％～14.60％.
Abstract
Currently,manual detection is the mainstream method for detecting the appearance damage of disassembled electric energy meters,which seriously restricts the recognition effi-ciency of disassembled electric energy meters.Therefore,a deep convolutional neural network model RAN based on attention mechanism and residual idea is proposed.According to the atten-tion mechanism,channel and spatial attention modules are constructed in RAN to enhance the feature extraction ability of the network,and residual idea is used to avoid the feature value at-tenuation caused by the attention module.The research objects include fire gauges,water im-mersion gauges,terminal abnormality gauges,display screen damage gauges,appearance rup-ture gauges,and normal gauges.The average recognition rate of RAN is 94.58％,which is 0.32％ to 14.60％ higher than ResNet101,VGG16,MobileNet,and ShuffleNet.
关键词
拆回电能表/外观破损/注意力机制/残差神经网络Key words
disassembled electric energy meter/Appearance damage/Attention mechanism/Re-sidual neural network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据