基于改进卷积神经网络算法的违规作业行为检测  

Illegal operation detection based on improved convolution neural network algorithm

作　　者：赵连斌 张锋 杨辉 ZHAO Lianbin;ZHANG Feng;YANG Hui(State Grid Gansu Electric Power Company,Lanzhou 730000,China;Dingxi Power Supply Company,State Grid Gansu Electric Power Company,Dingxi 743000,China)

机构地区：[1]国网甘肃省电力公司,甘肃兰州730000 [2]国网甘肃省电力公司定西供电公司,甘肃定西743000

出　　处：《电子设计工程》2023年第21期141-145,共5页Electronic Design Engineering

基　　金：国网甘肃省电力公司安全劳动技术保护科技兴安工程项目(SGGSDX00AJWT2101081)。

摘　　要：为了提升电力生产环境下违规作业行为的检测效率,文中对卷积神经网络的相关理论进行了研究,将二维平面下的卷积、池化运算扩展到了三维空间(C3D),使得网络在特征提取时可以有效获取视频帧信息。借鉴Inception网络的思路,使用更小颗粒的卷积结构替代C3D网络中的大颗粒卷积运算,有效提升了网络的感知能力和非线性拟合能力。此外,还对传统的随机梯度下降(SGD)训练方式进行了改进,引入了一种基于分数阶动量的梯度下降法,该方法使用训练动量进行自适应训练调节,有效解决了SGD训练误差不稳定、容易陷入局部最优等缺点。以某供电公司安监部门采集的视频数据集为样本进行的性能测试结果表明,其识别精度可达92.25%,相较于普通C3D网络,识别精度提升了4.89%,训练时间下降了61.41%。In order to improve the detection efficiency of illegal operation in power production environment,this paper studies the relevant theory of convolution neural network,and extends the convolution and pooling operation in two-dimensional plane to three-dimensional space(C3D),so that the network can effectively obtain video frame information in feature extraction.Using the idea of Inception network for reference,the convolution structure of smaller particles is used to replace the convolution operation of large particles in C3D network,which effectively improves the perception ability and nonlinear fitting ability of the network.In addition,the traditional Stochastic Gradient Descent(SGD) training method is improved,and a gradient descent method based on fractional momentum is introduced.This method uses the training momentum for adaptive training adjustment,which effectively solves the shortcomings of unstable SGD training error and easy to fall into local optimization.Taking the video data set collected by the safety supervision department of a power supply company as the sample,the performance test results show that the recognition accuracy can reach 92.25%.Compared with the ordinary C3D network,the recognition accuracy is improved by 4.89% and the training time is reduced by 61.41%.

关 键 词：卷积神经网络 C3D 分数阶动量 梯度下降 视频识别 违规判别 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]