摘要
电力金具上销钉松动、缺失等常见缺陷严重影响着电力系统的稳定运行,而该类缺陷的检测方式主要依赖于人工标注,致使效率低下,为此提出利用RetinaNet算法实现销钉缺陷智能识别的方法。该方法首先研究算法的主要参数——学习率对模型训练速度和销钉缺陷识别率的影响;继而考虑到销钉松动类数据样本较少以及该类样本收集代价较高的情况,通过添加辅助数据样本来缓解类别失衡造成的影响;最后为消除辅助数据与目标数据样本之间存在的差异而产生的影响,对相关辅助与目标数据进行了量化分析。实验结果表明,缺陷数据样本的不足使得训练好的模型容易将少数类错误识别为数量偏多的正常类别,通过添加一定数量的相关辅助数据样本,能够有效地缓解类别失衡带来的不利影响,完成销钉缺陷智能识别。
Common defects of power fittings such as pin looseness and missing seriously affect stable operation of the power system,while detection methods for these defects mainly depend on manual marking which may cause low efficiency.Thus,this paper proposes a method based on RetinaNet algorithm for intelligent identification on defects of pins.This method firstly studies effect of the main parameter learning rate of RetinaNet algorithm on training speed of the model and defect identification rate.Then considering less data samples of looseness category and high price of sample collection,this method relieves influence of category imbalance by adding auxiliary data samples.Finally,it makes quantitative analysis on related auxiliary and objective data so as to remove effect caused by differences between auxiliary data samples and objective data samples.The test result indicates insufficient defected data samples may cause the trained model easy to identify minority category errors as normal categories with more numbers.By adding a certain amount of related auxiliary data samples,it can effectively relieve adverse effect caused by category imbalance and complete intelligent identification on defects of pins.
作者
王凯
王健
刘刚
周文青
何卓阳
WANG Kai;WANG Jian;LIU Gang;ZHOU Wenqing;HE Zhuoyang(School of Electric Power,South China University of Technology,Guangzhou,Guangdong 510640,China;Guangdong Power Transmission and Transformation Engineering Co.,Ltd.,Guangzhou,Guangdong 510160,China)
出处
《广东电力》
2019年第9期41-48,共8页
Guangdong Electric Power
基金
国家高技术研究发展计划项目(863计划)(2015AA050201)
关键词
缺陷检测
深度学习
辅助数据
类别失衡
量化分析
defect detection
deep learning
auxiliary data
category imbalance
quantitative analysis
