迁移学习和卷积神经网络电力设备图像识别方法  

Power equipment image recognition method based on transfer learning and convolutional neural network

作　　者：王昕 赵飞[2] 蒋佐富 尚将 吴瑞文 WANG Xin;ZHAO Fei;JIANG Zuofu;SHANG Jiang;WU Ruiwen(Taizhou Hongchuang Power Group Limited Company,Taizhou 318000,China;School of Electric and Electrical Engineering,North China Electric Power University,Baoding 071000,China;State Grid Zhejiang Taizhou Huangyan Power Supply Co.,Ltd.,Taizhou 318000,China)

机构地区：[1]台州宏创电力集团有限公司,浙江台州318000 [2]华北电力大学电气与电子工程学院,河北保定071000 [3]国网浙江台州市黄岩区供电有限公司,浙江台州318000

出　　处：《中国测试》2020年第5期108-113,共6页China Measurement & Test

基　　金：国家自然科学基金项目(51177047)。

摘　　要：一线生产单位一般不具备建立大量电力设备图像数据集的条件,因此在使用深度学习模型协助完成对电力设备图像的识别过程中受到限制。通过对电力设备进行三维建模和多角度渲染,获得大量模拟电力设备图像,解决深度学习模型卷积神经网络在学习过程中数据集不足的问题。同时,通过迁移学习的方式将经过模拟电力设备图像训练的卷积神经网络应用于对真实电力设备图像的学习中,提高学习效率和精度,最终取得93.5%的识别准确率。该方法为一线生产单位将卷积神经网络应用于电力设备图像识别分类任务提供一种解决办法。Front-line production units generally do not have the conditions to establish a large number of power equipment image data sets,so they are limited in the use of deep learning model to assist in the process of power equipment image recognition.Through three-dimensional modeling and multi-angle rendering of power equipment,a large number of images of power equipment were obtained,which solved the problem of insufficient data set in the learning process of deep learning model convolutional neural network.At the same time,the convolution neural network trained by the simulated power equipment image is applied to the learning of real power equipment image by means of transfer learning,which improves the learning efficiency and accuracy and finally achieves the recognition accuracy of 93.5%.This method provides a solution for the firstline production units to apply the convolutional neural network to the power equipment image recognition and classification task.

关 键 词：图像识别 电力设备 迁移学习 卷积神经网络 

分 类 号：TM726[电气工程—电力系统及自动化]