| 引用本文: | 许伯强,吴咏诗,尹彦博,孙丽玲.基于跨范式特征融合与小样本学习的异步电机红外图像故障诊断[J].电力自动化设备,2026,46(1):202-208 |
| XU Boqiang,WU Yongshi,YIN Yanbo,SUN Liling.Infrared image-based fault diagnosis of asynchronous motor based on cross-paradigm feature fusion and few-shot learning[J].Electric Power Automation Equipment,2026,46(1):202-208 |
|
| 摘要: |
| 异步电机的红外图像故障诊断面临数据稀缺性和特征提取能力不足的挑战。为解决这一问题，提出了一种基于跨范式特征融合(CPFF)与小样本学习的故障诊断模型。该模型结合ConvNeXt提取局部特征与Swin Transformer提取全局特征，通过自空间自适应融合模块(SSAFM)实现高效特征融合。SSAFM利用自注意力和空间注意力机制进一步增强特征表达能力。模型在包含10种故障类别和空载状态的异步电机红外图像数据集上，以每类1张真实图像进行训练，并通过数据增强生成伪验证集优化超参数。实验结果表明，该模型在真实红外图像测试集上的分类精度可达到95.14 %，显著优于ConvNeXt、Swin Transformer及其他先进分类模型。该研究可为小样本条件下的异步电机红外图像故障诊断提供解决方案。 |
| 关键词: 异步电机 红外图像 故障诊断 特征融合 小样本学习 |
| DOI：10.16081/j.epae.202507001 |
| 分类号:TM307 |
| 基金项目:国家自然科学基金资助项目(51277077) |
|
| Infrared image-based fault diagnosis of asynchronous motor based on cross-paradigm feature fusion and few-shot learning |
|
XU Boqiang, WU Yongshi, YIN Yanbo, SUN Liling
|
|
Department of Electrical Engineering, North China Electric Power University, Baoding 071003, China
|
| Abstract: |
| Infrared image-based fault diagnosis of asynchronous motors faces significant challenges due to data scarcity and limited feature extraction capabilities. To address these issues, a fault diagnosis model based on cross-paradigm feature fusion(CPFF) and few-shot learning is proposed. The model integrates local features extracted by ConvNeXt and global features extracted by Swin Transformer, and achieves efficient feature fusion through a self spatial adaptive fusion module(SSAFM). SSAFM employs self-attention and spatial attention mechanisms to further enhance feature representation. The model is trained on the infrared image dataset of asynchronous motor with 10 fault categories and one no-load condition, using one real image of each category, and the pseudo-validation set is generated via data augmentation to optimize the hyperparameter. Experimental results demonstrate that the model achieves a classification accuracy of 95.14 % on a real infrared image test set, significantly outperforming ConvNeXt, Swin Transformer and other advanced classification models. This study presents a viable solution for fault diagnosis of asynchronous motors based on infrared imagery under few-shot learning conditions. |
| Key words: asynchronous motor infrared images fault diagnosis feature fusion few-shot learning |
