基于融合G-CNN的轻量化智能巡检识别算法设计  

Design of lightweight intelligent patrol recognition algorithm based on fusion G-CNN

作　　者：周强 李成玉 李永红 徐雪莹 ZHOU Qiang;LI Chengyu;LI Yonghong;XU Xueying(Shandong Yimeng Pumped Storage Co.,Ltd.,Linyi 273400,China)

机构地区：[1]山东沂蒙抽水蓄能有限公司,山东临沂273400

出　　处：《电子设计工程》2024年第24期81-85,共5页Electronic Design Engineering

基　　金：山东省沂蒙抽水蓄能电站项目(SGXYYMOOJDMM2000055)。

摘　　要：为了提升抽水蓄能发电站内巡检机器人的检测精度和效率,文中对深度学习框架下的图像识别算法展开研究。对传统卷积神经网络(CNN)进行了改进,提出一种基于Gabor CNN网络的电力智能巡检算法(G-CNN),该算法结合了Gabor滤波器的优势,具有更好的方向选择性和尺度选择性。基于抽水蓄能发电站内巡检机器人采集到的图像构建了数据集,并以此对所设计的网络进行训练,训练后的网络能够有效地对电力设备进行检测和缺陷识别。对所提算法进行的性能对比测试结果表明,引入Gabor层后,G-CNN网络较同结构的CNN网络在识别精度上提升了4.16%,同时运行时间缩短了56.59%,可以满足智能巡检的轻量化需求。In order to improve the detection accuracy and efficiency of inspection robots in pumped storage power plants,this paper conducts research on image recognition algorithms under the deep learning framework.An improvement has been made to the traditional Convolutional Neural Network(CNN),proposing a Gabor CNN(G-CNN)based intelligent power inspection algorithm.This algorithm combines the advantages of Gabor filters and has better directional and scale selectivity.A dataset was constructed based on the images collected by the inspection robot inside the pumped storage power station,and the designed network was trained accordingly.The trained network can effectively detect and identify defects in power equipment.The performance comparison test results of the proposed algorithm show that after introducing the Gabor layer,the G-CNN network has improved recognition accuracy by 4.16%compared to CNN networks of the same structure,while reducing runtime by 56.59%,which can meet the lightweight requirements of intelligent inspection.

关 键 词：抽水蓄能 电力巡检 图像识别 卷积神经网络 深度学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论] TN912.34[自动化与计算机技术—计算机科学与技术]