摘要
针对电缆局部放电灰度图数目过少,难以训练基于大规模数据集的深度残差网络模型的问题,从数据增广与模型简化两方面入手,提出一种基于少样本的电缆局部放电模式识别方法。选择深度卷积生成对抗网络(DCGAN)对局部放电灰度图进行扩充,构造基于盒维数的生成样本评价指标以验证生成样本的有效性;对比分析残差模块、网络深度对深度残差网络(ResNet)分类性能的影响,在此基础上提出一种简化的残差网络模型来匹配小规模局部放电数据集。经实验测试,简化后的残差网络模型平均迭代时间为7.3s,识别准确率达98.5%。与直接使用少量样本训练深度残差网络的方法相比,所提方法具有较快的模型训练速度与较高的识别精度。
Aiming at the problem that the number of partial discharge gray-scale images for power cables is small, and it is difficult to train deep residual network model based on large-scale data sets. A method of partial discharge pattern recognition for cables under few samples condition is proposed. The method uses the idea of combining expanded samples with simplified model. A data augmentation method based on deep convolutional generative adversarial network is proposed for the partial discharge grayscale image. In order to verify the effectiveness of the generated samples, an evaluation index of the generated samples based on the box dimension is proposed. The effects of residual modules and network depth on the classification performance of residual networks are compared. A simplified residual network model is proposed to match the small-scale partial discharge data set. The simplified residual network model is tested by experiments. The average iteration time of the network model is 7.3s, and the recognition accuracy rate reaches 98.5%. Compared with the method of directly using few samples to train the deep residual network, the proposed method has faster model training speed and higher recognition accuracy.
作者
岳云飞
孙抗
YUE Yun-fei;SUN Kang(School of Electrical Engineering and Automation,Henan Polytechnic University;Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment,Jiaozuo 454003,China)
出处
《软件导刊》
2023年第2期60-66,共7页
Software Guide
基金
河南省科技攻关项目(202102210092)
河南省高等学校青年骨干教师培养计划项目(2021GGJS056)。
关键词
生成对抗网络
残差网络
局部放电
模式识别
深度学习
generative adversarial network
residual network
partial discharge
pattern recognition
deep learning
