Multi-scale information fusion model for feature extraction of converter transformer vibration signal
Multi-scale information fusion model for feature extraction of converter transformer vibration signal
复制标题
DOI：
10.1016/j.measurement.2021.109555
复制
发表时间：
2021-08
期刊：
影响因子：
5.6
通讯作者：
Rui Xiao;Zhanlong Zhang;Yongye Wu;Peiyu Jiang;Jun Deng
中科院分区：
文献类型：
--
作者：
Rui Xiao;Zhanlong Zhang;Yongye Wu;Peiyu Jiang;Jun Deng
换流变压器振动信号特征提取模型存在精度不高的问题。主要原因是训练过程中容易丢失与时间相关的信息。近年来，卷积神经网络(CNN)被证明打破了传统分类模型的局限性。美国有线电视新闻网在大规模图像识别任务中取得了显著成效。针对换流变压器振动信号建模困难、时间相关信息丢失的问题，提出了一种换流变压器振动信号多尺度融合特征提取模型。该模型包括图像生成模块和多尺度信息融合模块。图像生成模块通过计算马尔可夫变换场和连续小波变换，将振动信号转换为时域、频域和能量特征图像。在多尺度信息融合模块中，将注意力模块引入到多并联卷积神经网络中，对特征图像进行融合。该方法可以充分利用深度学习的优势，生成能够从多尺度上表示时间序列的图像。为了评价模型的有效性，我们建立了一个基于实测振动信号的数据集，并通过分类实验进行了检验。结果表明，本文的模型状态识别准确率为96.15%，优于LSTM和1D-CNN等经典的时间序列处理网络。
The vibration signal feature extraction model of the converter transformer has the problem of low accuracy. The main reason is that the time-related information is easily lost during training. In recent years, convolutional neural networks (CNN) have been proven to break the limits of traditional classification models. CNN has achieved remarkable results in large-scale image recognition tasks. To solve the problem of difficulty in model building and loss of time-related information, this paper proposes a multi-scale fusion feature extraction model for converter transformer vibration signal. The model includes an image generation module and a multi-scale information fusion module. The image generation module converts vibration signals into time-domain, frequency-domain, and energy feature images by calculating the Markov Transform Field and Continuous Wavelet Transform. In the multi-scale information fusion module, the attention module is introduced into the multi-parallel convolutional neural network to fuse the feature images. The proposed method can make full use of the advantages of deep learning through generating images that can represent time series from multi-scale. To evaluate the effectiveness of the model in this paper, we establish a data set based on the measured vibration signals and tested by classification experiments. The result shows the accuracy of the model state recognition in this paper is 96.15%, which is better than classic time series processing networks such as LSTM and 1D-CNN.