| 王付安*,占天行**,赵媛媛*,冯坤**,吴斯琪*,杨雨琦**.时频图像多模态的电机故障特征提取融合算法[J].高技术通讯(中文),2025,35(12):1364~1374 |
| 时频图像多模态的电机故障特征提取融合算法 |
| Motor fault feature extraction fusion algorithm based on multimodal time-frequency images |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 12. 009 |
| 中文关键词: 故障诊断； 多模态； 特征提取； 卷积神经网络； 小波变换 |
| 英文关键词: fault diagnosis, multimodal, feature extraction, convolutional neural network, wavelet transform |
| 基金项目: |
| 作者 | 单位 | | 王付安* | （* 中国五洲工程设计集团有限公司北京100053）
（** 北京化工大学高端压缩机及系统技术全国重点实验室北京 100029） | | 占天行** | | | 赵媛媛* | | | 冯坤** | | | 吴斯琪* | | | 杨雨琦** | |
|
| 摘要点击次数: 35 |
| 全文下载次数: 28 |
| 中文摘要: |
| 传统的单模态检测方法容易受到噪声干扰，而多模态方法能够有效融合来自不同模态的信息特征。本文提出了一种基于时频图像的多模态电机故障特征提取与融合算法。该方法收集三相交流异步电机的振动、声音和电流3个模态的数据，并从时频信号中提取特征。特征提取包括2种方法：一是通过时频分析生成图像样本，使用深度残差网络（residual network，ResNet）进行图像特征提取；二是使用小波变换提取时频信号的局部信息，从而去除由环境等因素引起的高频噪声成分。将这两部分特征输入到自编码器中进行降维和去冗余处理，并对不同特征进行融合。最后，使用支持向量机分类器对融合后的特征进行分类，评估其分类准确率。实验结果表明，与单模态分类网络相比，本文方法的分类准确率更高，验证了该方法的有效性与可行性。 |
| 英文摘要: |
| Traditional unimodal detection methods are susceptible to noise interference, whereas multimodal approaches can effectively integrate information from different modalities. This paper proposes a multimodal motor fault feature extraction and fusion algorithm based on time-frequency images. The proposed method collects data from three modalities—vibration, sound, and current—of a three-phase asynchronous motor and extracts features from the corresponding time-frequency signals. Feature extraction involves two approaches: first, time-frequency analysis is used to generate image samples, and a deep residual network (ResNet) is employed for image feature extraction; second, wavelet transform is applied to extract localized information from the time-frequency signals, removing high frequency noise components caused by environmental factors. The two sets of features are then fed into an autoencoder for dimensionality reduction and redundancy removal, ultimately achieving feature fusion. Finally, a support vector machine classifier is used to classify the fused features, and assess classification accuracy. Experimental results demonstrate that, compared to unimodal classification networks, the proposed method achieves higher classification accuracy, supporting the effectiveness and feasibility of the approach. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
