电工技术学报  2024, Vol. 39 Issue (20): 6577-6590    DOI: 10.19595/j.cnki.1000-6753.tces.231335
高电压与放电 |
基于变权属性矩阵的变压器零样本故障诊断技术
雷蕾潇1, 何怡刚1, 姚其新2, 邢致恺1
1.武汉大学电气与自动化学院 武汉 430072;
2.国网湖北省电力有限公司直流公司 宜昌 443000
Zero-Shot Fault Diagnosis Technique of Transformer Based on Weighted Attribute Matrix
Lei Leixiao1, He Yigang1, Yao Qixin2, Xing Zhikai1
1. School of Electrical Engineering and Automation Wuhan University Wuhan 430072 China;
2. State Grid Hubei Direct Current Company Yichang 443000 China
全文: PDF (2531 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 针对变压器故障数据的稀缺性及数据分布中存在的长尾现象,导致故障诊断准确率低的问题,该文提出一种基于变权属性矩阵的变压器零样本故障诊断技术。首先,采用改进的高效通道注意力网络-堆栈式自编码器(IECANet-SAE)网络构建特征提取模块,自适应地提取油中气体数据的关键特征信息;其次,利用基于潜在狄利克雷分布的主题建模方法构建变权属性矩阵;最后,提出基于神经网络的朴素贝叶斯(NNB)方法学习已知故障特征信息与属性矩阵的空间映射关系,建立零样本故障诊断模型并依靠模型实现未知故障类型诊断。应用IEC TC 10故障数据库及典型故障数据对所提方法加以验证。试验结果表明,所提出的方法具有更好的诊断效果,且在零样本条件下故障诊断平均准确率高达83%,平均诊断时间达0.18 s。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
雷蕾潇
何怡刚
姚其新
邢致恺
关键词 零样本学习特征提取故障诊断变权属性矩阵    
Abstract:In actual engineering applications, transformer failure is a small probability event, so there is little fault data and low data quality, and the actual data distribution often exhibits a long tail effect. However, traditional data-driven fault diagnosis methods require a large amount of fault data for sample training, which leads to problems such as low model accuracy and model failure in the actual application of traditional supervised models. To address this type of problem, this article proposes a zero-shot fault diagnosis technology for transformers based on variable weight attribute matrix. The spatial mapping function between known fault attribute matrices and dissolved gas data is learned to achieve accurate identification of unknown faults.
Firstly, an improved efficient channel attention network-stack autoencoder (IECANet-SAE) network is used to construct a data feature extraction network. Among them, convolutional neural networks extract local important information from input data through convolutional operations, while efficient channel attention networks focus on important information by learning attention weights between channels. The use of feature extraction networks solves the problem of quality and quantity of fault sample data, while adaptively extracting key feature information from dissolved gas data. Secondly, based on the description information of transformer faults, a variable weight attribute matrix is constructed using a topic modeling method based on potential Dirichlet distribution. This includes fault type classification, topic word extraction, similarity calculation, and matrix construction. By enhancing the connection between fault labels and fault information through the fault attribute matrix, data sharing between known and unknown class faults can be achieved. Then, the neural network-based Naive Bayes (NNB) method is used to learn the mapping function between each attribute vector in the fault attribute matrix and the important feature information of known class faults in the feature space. Based on the mapping function, the unknown class feature information is mapped to the same feature space to obtain the corresponding attribute vector. Finally, based on Bayesian principle, the maximum likelihood estimation formula is used to obtain the most likely fault type for unknown class faults.
The proposed method was validated using the IEC TC 10 fault database and relevant authoritative literature that has been publicly published. From the accuracy of fault diagnosis results, it can be seen that data-driven fault diagnosis models are difficult to achieve fault diagnosis without sample training, so the diagnostic accuracy of this type of algorithm is 0. On the basis of the transformer fault attribute matrix proposed in this article, zero-shot fault diagnosis is achieved, which has a higher accuracy of up to 0.83 compared to direct attribute prediction (DAP) and indirect attribute prediction (IAP) methods. From the fault diagnosis time, it can be seen that the DAP and IAP methods analyze the original data, so the calculation time is slightly longer due to the large amount of data. The method in this article retains some important features in the data, reducing computational time to a certain extent. The diagnostic time is about 0.18 seconds, which is slightly faster than other methods.
The following conclusions can be drawn through experimental analysis: (1) The use of LDA based topic modeling method effectively enhances the correlation between fault information and fault sample labels, solving the problem of difficult fault diagnosis under the condition of no training samples. (2) Extracting the main features of dissolved gas data through the IECANet-SAE network has better fault feature separation performance compared to other methods, which helps to improve transformer fault diagnosis performance. (3) The NNB algorithm is used to construct the mapping relationship between important feature data and fault attribute matrix, efficiently and accurately identifying unknown fault types. The average accuracy and diagnostic time of the diagnostic results are superior to other methods.
Key wordsZero-shot learning    feature extraction    fault diagnosis    weighted attribute matrix   
收稿日期: 2023-08-17     
PACS: TM407  
基金资助:国家重点研发计划“储能与智能电网技术”专项“海上风电并网系统远程监测与故障诊断技术”资助项目(2023YFB2406900)
通讯作者: 何怡刚 男,1966年生,教授,博士生导师,研究方向为混合信号电路故障诊断、电子设备可靠性和通信信道建模与监测等。E-mail: yghe1221@whu.edu.cn   
作者简介: 雷蕾潇 女,1994年生,博士研究生,研究方向为变压器故障诊断。E-mail: leixiaolei@whu.edu.cn
引用本文:   
雷蕾潇, 何怡刚, 姚其新, 邢致恺. 基于变权属性矩阵的变压器零样本故障诊断技术[J]. 电工技术学报, 2024, 39(20): 6577-6590. Lei Leixiao, He Yigang, Yao Qixin, Xing Zhikai. Zero-Shot Fault Diagnosis Technique of Transformer Based on Weighted Attribute Matrix. Transactions of China Electrotechnical Society, 2024, 39(20): 6577-6590.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.231335          https://dgjsxb.ces-transaction.com/CN/Y2024/V39/I20/6577