一种用于变压器故障诊断的卷积神经网络优化方法

doi:10.16180/j.cnki.issn1007-7820.2023.12.011

摘要/Abstract

摘要：

传统故障诊断方法存在编码不完整和编码边界过于绝对等缺点,难以满足电网运维工作的实际需求。利用电力变压器故障时产生的气体对变压器进行故障诊断是目前智能电网状态检测的热门研究领域。但变压器出现各类故障的频率具有较大差别,可能造成故障样本信息不完整、数据量不足,使用传统卷积神经网络模型会出现训练过程收敛速度慢、训练精度低等问题。文中提出一种在保留原始数据特征不变的情况下进行数据增强的方法,将扩充后的一维数据转化为二维图片输入到二维卷积神经网络诊断模型中,并对卷积神经网络模型中的Adam优化算法进行改进。诊断结果表明,所提方法使得网络训练的精度达到了96.20%,相较于传统的一维卷积神经网络故障诊断方法(92.12%)具有更高的收敛速度和泛化能力。

关键词: 电力变压器, DGA, 故障诊断, 机器学习, 数据增强, 二维卷积, 神经网络, Adam

Abstract:

Traditional fault diagnosis methods have disadvantages such as incomplete coding and overly absolute coding boundaries, which are difficult to meet the actual needs of power grid operation and maintenance. Using the gas generated when a power transformer fails to diagnose the transformer fault is currently a popular research area for smart grid condition detection. However, the frequency of various types of faults in transformers varies greatly, which may result in incomplete fault sample information and insufficient data, and the traditional convolutional neural network models have problems such as unstable training process, low training accuracy and long time. Based on the transformer fault diagnosis technology of one dimensional convolutional neural network, this study proposes a new method of data enhancement while keeping the original data features unchanged, transforming expanded one dimensional data into two dimensional pictures to input into the two dimensional convolutional neural network diagnosis model, and improve the Adam optimization algorithm in the convolutional neural network model architecture. Diagnostic results indicate that the accuracy of network training reaches 96.20%. At the same time, it has higher convergence speed and generalization ability than the traditional one dimensional convolutional neural network fault diagnosis method(92.12%).

Key words: power transformer, DGA, fault diagnosis, machine learning, data enhancement, two dimensional, neural network, Adam

中图分类号:

TM41

汪徐阳,易映萍,李田丰. 一种用于变压器故障诊断的卷积神经网络优化方法[J]. 电子科技, 2023, 36(12): 79-86.

WANG Xuyang,YI Yingping,LI Tianfeng. A Convolutional Neural Network Optimization Method for Fault Diagnosis of Power Transformer[J]. Electronic Science and Technology, 2023, 36(12): 79-86.

图/表 13

图1

图2

图3

表1

表2

表3

表4

表5

表6

图4

图5

表7

图6

参考文献 18

[1]	代杰杰. 基于深度学习的变压器状态评估技术研究[D]. 上海: 上海交通大学, 2018:2-4.
	Dai Jiejie. Research on transformer state assessment based on deep learning[D]. Shanghai: Shanghai Jiao Tong University, 2018:2-4.
[2]	邓焱. 基于多维信息融合的变压器状态评估与故障诊断方法研究[D]. 广州: 华南理工大学, 2018:4-10.
	Deng Yan. A study on transformer state evaluation and fault diagnosis based on multi-dimensional information fusion[D]. Guangzhou: South China Universityof Technology, 2018:4-10.
[3]	Gao Z W, Cecati C, Ding S X. A survey of fault diagnosis and fault-tolerant techniques—Part I:Fault diagnosis with model-based and signal-based approaches[J]. IEEE Transactions on Industrial Electronics, 2015, 62(6):3757-3767. doi: 10.1109/TIE.2015.2417501
[4]	Zhang X, Zhu M Z. Research on transformer fault diagnosis method based on rough set optimization BP neural network[C]. Beijing: The Sencond IEEE Conference on Energy Internet and Energy System Integration, 2018:1-5.
[5]	尹金良, 刘玲玲. 代价敏感相关向量机的研究及其在变压器故障诊断中的应用[J]. 电力自动化设备, 2014, 34(5):111-115.
	Yin Jinliang, Liu Lingling. CS-RVM and its application in fault diagnosis of power transformers[J]. Electric Power Automation Equipment, 2014, 34(5):111-115.
[6]	宋志杰, 王健. 模糊聚类和LM算法改进BP神经网络的变压器故障诊断[J]. 高压电器, 2013, 55(5):54-59.
	Song Zhijie, Wang Jian. Transformer fault diagnosis based on BP neural network optimized by fuzzy clustering and LM algorithm[J]. High Voltage Apparatus, 2013, 55(5):54-59.
[7]	王德文, 雷倩. 基于贝叶斯正则化深度信念网络的电力变压器故障诊断方法[J]. 电力自动化设备, 2018, 38(5):129-135.
	Wang Dewen, Lei Qian. Fault diagnosis of power transformer based on BR-DBN[J]. Electric Power Automation Equipment, 2018, 38(5):129-135.
[8]	许静. 基于受限玻尔兹曼机的变压器故障诊断[D]. 北京: 华北电力大学, 2017:12-13.
	Xu Jing. Transformer fault diagnosis using restricted boltzmann machine[D]. Beijing: North China Electric Power University, 2017:12-13.
[9]	Zhang L J, Sheng G H, Hou H J, et al. A fault diagnosis method of power transformer based on cost sensitive one-dimensional convolution neural network[C]. Chengdu: The Fifth Asia Conference on Power and Electrical Engineering, 2020:1824-1828.
[10]	Lin N, Guo Z W. Transformer fault diagnosis model based on FI-CNN method[C]. Zhengzhou: International Conference on Intelligent Traffic Systems and Smart City, 2022:9-16.
[11]	Ma L C, Zhang Y, Sun P, et al. Research on fault diagnosis and optimization of crusher based on atom search algorithm-BP neural network[C]. Hefei: Chinese Control And Decision Conference, 2020:671-676.
[12]	Bai J Y, Zhang J W. BGADAM:Boosting based genetic-evolutionary ADAM for neural network optimization[C]. Qingdao: International Joint Conference on Neural Networks, 2021:1-8.
[13]	Lai G X, Li F P, Feng J Q. A LPSO-SGD algorithm for the optimization of convolutional neural network[C]. Wellington: IEEE Congress on Evolutionary Computation, 2019:1038-1043.
[14]	Loshchilov I, Hinton G, Krizhevsky A. Dropout:A simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014, 15(1):1929-1958.
[15]	Zhang Z J. Improved adam optimizer for deep neuralnetworks[C]. Banff: IEEE/ACM the Twenty-sixth International Symposium on Quality of Service, 2018:1-2.
[16]	徐康健, 汪卫国. GB/T7252—2001《变压器油中溶解气体分析和判断导则》中一问题的探讨[J]. 电力技术, 2009, 15(2):62-63.
	Xu Kangjian, Wang Weiguo. GB/T7252—2001 Guidelines for the analysis and judgment of dissolved gasesin transformer oil[J]. Electric Power Technology, 2009, 15(2):62-63.
[17]	尹金良. 基于相关向量机的油浸式电力变压器故障诊断方法研究[D]. 北京: 华北电力大学, 2013:7-10.
	Yin Jinliang. Research on oil-immersed power transformer fault diagnosis based on relevance vector machine[D]. Beijing: North China Electric Power University, 2013:7-10.
[18]	李志恒, 何军, 胡昭华. 基于dropout正则化的半监督域自适应方法[J]. 计算机应用研究, 2021, 38(2):591-594,599.
	Li Zhiheng, He Jun, Hu Zhaohua. Semi-supervised dom-ain adaptive method based on dropout regularization[J]. Application Research of Computers, 2021, 38(2):591-594,599.

编号	网络层	卷积核大小	卷积核数目
C₁	卷积层	3×3	32
S₂	最大池化层1	2×2	32
C₃	卷积层2	1×1	64
S₄	最大池化层2	2×2	64
C₅	卷积层3	2×2	100
F₆	全连接层	100	1
L₇	输出层	6	1

气体含量/μL·L^-1					故障类型
H₂	CH₄	C₂H₆	C₂H₄	C₂H₂	故障类型
110.00	5.000	2.700	3.200	1.7	正常
14.67	3.680	10.540	2.710	0.2	低能放电
65.00	26.100	10.100	41.600	57.8	高能放电
218.00	120.700	126.600	142.100	0.9	中低温过热
26.50	298.000	0.000	69.300	8.6	高温过热
180.85	0.574	0.234	0.188	0.0	局部放电

编号	故障类型	编码
1	正常	000001
2	低能放电故障	000010
3	高能放电故障	000100
4	中低温过热故障	001000
5	高温过热故障	010000
6	局部放电故障	100000

模型	测试集诊断准确率/%			平均诊断准确率/%	训练时间/s
模型	T₁	T₂	T₃	平均诊断准确率/%	训练时间/s
1D CNN	90.37	91.94	92.12	91.48	266.78
2D CNN	93.26	95.38	96.20	94.94	211.38

场景	S₁	S₂
训练集精确度	95.37	97.22
测试集精确度	90.50	95.28