摘要
高频变压器是电力电子变压器的核心元件。为提升基于油色谱可视化的高频变压器故障识别方法的准确率,提出了一种基于递归图和MobileNetV3的变压器故障识别方法。首先,对于一维油色谱故障样本序列利用递归图算法进行可视化处理,绘制出故障样本二维真彩图;其次,对于故障样本图片较少无法满足深度学习需求的问题,利用深度卷积生成对抗网络进行样本扩充;最后,基于扩充后的样本集利用MobileNetV3模型进行训练。试验结果表明:相较于其余可视化方法,递归图转换的真彩图的区分度更高;深度卷积生成对抗网络生成的样本集更加符合真实情况。改进的MobileNetV3模型可以在故障种类复杂的情况下对故障进行精确识别,且算法收敛速度快、模型小,更加适用于实际应用。
High frequency(HF)transformer is the core component of power electronic transformer.To improve the accuracy of fault identification method for HF transformers based on oil chromatography visualization,a transformer fault identification method based on recurrence plots and MobileNetv3 was proposed.Firstly,the one-dimensional oil chromatography fault sample sequence was visualized using recurrence plots algorithm to draw a two-dimensional true color map of the fault sample;secondly,for the problem of insufficient fault sample images that could not meet the requirements of deep learning,a deep convolutional generative adversarial network was used for sample expansion;finally,based on the expanded sample set,the MobileNetv3 model was employed for training.The experimental results show that compared to other visualization methods,the true color images transformed from recurrence plots has higher discriminability;the sample set generated by deep convolutional generative adversarial networks is more in line with the real situation.The improved MobileNetV3 model can accurately identify faults in the case of complex faults,with fast algorithm convergence and small model size,making it more suitable for practical applications.
作者
陈超杰
顾华
沈晓峰
杨欢红
叶婧元
王宇轩
Chen Chaoje;Gu Hua;Shen Xiaofeng;Yang Huanhong;Ye Jingyuan;Wang Yuxuan(Qingpu Power Supply Branch,State Grid Shanghai Electric Power Co.,Ltd.,Shanghai 201700,China;School of Electrical Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《电气自动化》
2025年第3期1-5,共5页
Electrical Automation
基金
国网上海市电力公司科技项目(520934230008)。
关键词
高频变压器
变压器故障识别
油色谱
递归图
深度卷积生成对抗网络
high frequency(HF)transformer
transformer fault identification
oil chromatography
recurrence plots
deep convolutional generative adversarial network
