声纹信号-图形差分场增强和多头自注意力机制的变压器工作状态辨识方法
Transformer working state identification method based on voiceprint signal-motif difference field enhancement and multi-head self-attention mechanism
张寒 1熊云 1唐信 1王枭2
作者信息
- 1. 国网湖南超高压变电公司 长沙 410000;变电智能运检国网湖南省电力有限公司实验室 长沙 410000
- 2. 上海睿深电子科技有限公司 上海 200237
- 折叠
摘要
为提升电力变压器工作状态的智能监测水平,提出声纹信号-图形差分场增强和多头自注意力机制的变压器工作状态辨识方法.基于图形差分场技术将声纹信号映射为二维图像,再借助多头注意力机制的视觉转换器实现图像信息的深层挖掘与状态辨识,采用梯度加权类激活映射实现分类结果的可解释性分析.搭建了包含变压器4种典型工作状态下的实验模拟测试系统平台,实验结果表明:所提方法不仅能够有效表征变压器声纹信号的状态特征,且分类辨识精度相较于"时频图+引入多头注意力机制的变换网络"与"图形差分场+引入残差模块的卷积神经网络"的常规方法有显著提升,提升约6％,同时也具备较好的鲁棒性,可为电气设备的故障检测研究提供一定参考.
Abstract
In order to improve the intelligent monitoring level of power transformer working state,a method of transformer working state identification based on motif difference field(MDF)voiceprint signal enhancement and multi-head self-attention mechanism is proposed.Based on the MDF technology,the sound signal is mapped into a two-dimensional image,and then the depth mining and state recognition of image information are realized with the help of the visual converter of multi-head attention mechanism,and the explanatory power analysis of classification results is realized by using gradient weighted class activation mapping.The experimental simulation test system platform containing four typical operating states of the transformer is constructed,and the experimental results show that the proposed method not only can effectively characterize the state characteristics of the transformer acoustic signal,but also has a higher classification accuracy compared with the"time-frequency+transformer network with multi-head self-attention mechanism"and the"MDF+visual converter with multi-head attention mechanism",which is about 6％,and also has better robustness,which can provide a certain reference for the research on the detection of faults in electrical equipment.
关键词
图形差分场/多头自注意力机制/变压器/状态辨识Key words
Motif difference field/Multi-head self-attention mechanism/Transformer/State identification引用本文复制引用
基金项目
国家电网湖南省电力公司科技项目(5216A3210014)
出版年
2024
NETL
NSTL
维普
万方数据