摘要
传统的电能质量扰动识别在提取分类时都是人工手动单一地选择特征,这种方法普遍存在精度低、成本高、泛化能力弱等缺陷。针对目前深度学习在故障识别分类处理中的应用,文中提出一种基于多尺度融合选择卷积网络实现端到端的电能质量扰动识别分类方法。文中通过仿真实现了单一扰动信号和复合扰动信号,将Multi⁃scale目标检测和轻量级模块SKNet算法思路相结合,利用不同卷积核构造多个尺度下不同的特征,将特征融合选择得到新的主要特征量,并用线性分类器进行快速的扰动识别。研究结果表明,与未改进的深度学习识别方法对比,文中方法在特征提取上具有更强的区分空间、更高的识别率和鲁棒性,整体识别率达到99%,可为电能质量扰动识别研究提供一种新的思路。
In the traditional power quality disturbance identification,the features are selected manually and singly when extracting and classifying,which has many defects,such as low precision,high cost,and weak generalization ability.In allusion to the current application of deep learning in fault recognition and classification processing,an end⁃to⁃end power quality disturbance recognition and classification method based on multi⁃scale fusion selection convolutional network is proposed.The single disturbance signal and composite disturbance signal are realized by the simulation.The multi⁃scale target detection is combined with the idea of lightweight module SKNet algorithm.The different convolution kernels are used to construct the different features at multiple scales.The fusion selection of features are carried out to obtain new main features.And the linear classifier is used for the fast disturbance recognition.The research results show that,in comparison with the unimproved deep learning recognition methods,the proposed method has stronger discrimination space,higher recognition rate and robustness in the feature extraction,and its overall recognition rate can reach 99%,which can provide a new idea for power quality disturbance recognition.
作者
王凯
孙贤明
任成昊
胡玉耀
王文成
路尧
WANG Kai;SUN Xianming;REN Chenghao;HU Yuyao;WANG Wencheng;LU Yao(College of Electrical and Electronic Engineering,Shandong University of Technology,Zibo 255049,China;Zibo Wayouchuangfei Intelligent Technology Co.,Ltd.,Zibo 255000,China)
出处
《现代电子技术》
2022年第4期107-112,共6页
Modern Electronics Technique
基金
国家自然科学基金项目(51907109)。
关键词
电能质量扰动
扰动识别
特征提取
多尺度融合
卷积网络
融合选择
power quality disturbance
disturbance recognition
feature extraction
multi⁃scale fusion
convolutional network
fusion selection