| 引用本文: | 吴晓欣,何怡刚,段嘉珺,张慧,曾昭瑢.考虑复杂时序关联特性的Bi-LSTM变压器DGA故障诊断方法[J].电力自动化设备,2020,40(8): |
| WU Xiaoxin,HE Yigang,DUAN Jiajun,ZHANG Hui,ZENG Zhaorong.Bi-LSTM-based transformer fault diagnosis method based on DGA considering complex correlation characteristics of time sequence[J].Electric Power Automation Equipment,2020,40(8): |
|
| 摘要: |
| 当前基于油中溶解气体分析(DGA)的变压器故障诊断方法往往仅考虑单一时刻数据点，容错性差，难以充分挖掘在线监测数据的时序信息。提出一种考虑变压器油特征参量序列间复杂关系的基于双向长短时记忆(Bi-LSTM)网络的变压器故障诊断方法。首先构建了变压器油特征参量序列，基于序列数据构建了Bi-LSTM变压器故障诊断模型。工程实际中不同变压器油特征参量序列长短不一，需通过排序、分组填充对模型输入进行重构改进，然后对超参数进行优化。基于同一自建数据库对比所提方法与其他方法，结果表明：经过数据重构后所提方法的准确率可达91.9 %；当特征指标数量减少约2/3时，所提方法的准确率仅下降约1%，而其他方法的准确率平均下降约6 %；当采样数据存在10%的随机错误时，所提方法诊断准确率仅下降2%~6 %，且通过改变隐藏层的数量可得到改善。 |
| 关键词: 变压器 故障诊断 双向长短时记忆网络 时间序列 复杂关联关系 油中溶解气体分析 |
| DOI：10.16081/j.epae.202006004 |
| 分类号:TM41 |
| 基金项目:国家自然科学基金资助项目(51977153,51977161,51577046)；国家自然科学基金重点资助项目(51637004)；国家重点研发计划“重大科学仪器设备开发”资助项目(2016YFF0102200)；装备预先研究重点项目(41402040301) |
|
| Bi-LSTM-based transformer fault diagnosis method based on DGA considering complex correlation characteristics of time sequence |
|
WU Xiaoxin, HE Yigang, DUAN Jiajun, ZHANG Hui, ZENG Zhaorong
|
|
School of Electrical Engineering and Automation, Wuhan University, Wuhan 430072, China
|
| Abstract: |
| Transformer fault diagnosis methods based on DGA(Dissolved Gas Analysis) are commonly just consider data obtained at a single moment, which is susceptible to data errors and cannot fully mine the online monitoring data sequence information. Aiming at this problem, a Bi-LSTM(Bi-directional Long Short-Term Memory) network based transformer fault diagnosis method considering the complex correlation relationship among the characteristic parameter sequences of transformer oil is proposed. Firstly, the characteristic parameter sequences of transformer oil are constructed, and then the Bi-LSTM-based transformer fault diagnosis model is constructed based on the sequence data. Considering the different length of the sequences in engineering practice, the model inputs are reconstructed by sorting and group filling. The hyper-parameters are then optimized. Based on the same built database, the proposed method is compared with other methods. The results show that the diagnosis accuracy of the proposed method after data reconstruction can reach 91.9 %,and only drops by about 1% when the number of characteristic parameter decreased by about two-thirds, while other methods decrease by about 6 % on average. When the sampling data exists a 10% random errors, the diagnosis accuracy of the proposed method decreases by 2% to 6 %,and can be improved by changing the number of hidden layers. |
| Key words: electric transformers fault diagnosis bi-directional long short-term memory network time sequence complex correlation relationship dissolved gas analysis |
