Scholarlink

<p>To address the limitations of stator fault diagnosis methods for PMSMs that rely solely on single-phase current signals and therefore suffer from incomplete information, this work develops an innovative deep learning framework employing a multimodal cross-attention mechanism. First, a multi-scale dilated convolution and channel-attention-based (DC-CAM) residual block is introduced to extract discriminative local and long-range temporal features from the three-phase current and raw vibration signals, while adaptively enhancing informative channels through attention-driven feature recalibration. Subsequently, a modality fusion attention mechanism is introduced to integrate vibration features into each of the three-phase current channels. Finally, a three-phase cross-attention mechanism enables interaction and fusion among the current signals embedded with vibration information, thereby achieving end-to-end stator fault diagnosis for PMSMs. The method is validated on PMSM stator fault datasets with three different power ratings. The experimental results show that the multimodal cross-attention mechanism model proposed in this paper has higher fault diagnosis performance than other advanced models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A deep learning model with multimodal cross-attention mechanism for PMSM fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Jingna Liu,
  • Luchao Guo,
  • Cheng Wen,
  • Xiang Yao,
  • Zilei Duan,
  • Ruiting Zhang

摘要

To address the limitations of stator fault diagnosis methods for PMSMs that rely solely on single-phase current signals and therefore suffer from incomplete information, this work develops an innovative deep learning framework employing a multimodal cross-attention mechanism. First, a multi-scale dilated convolution and channel-attention-based (DC-CAM) residual block is introduced to extract discriminative local and long-range temporal features from the three-phase current and raw vibration signals, while adaptively enhancing informative channels through attention-driven feature recalibration. Subsequently, a modality fusion attention mechanism is introduced to integrate vibration features into each of the three-phase current channels. Finally, a three-phase cross-attention mechanism enables interaction and fusion among the current signals embedded with vibration information, thereby achieving end-to-end stator fault diagnosis for PMSMs. The method is validated on PMSM stator fault datasets with three different power ratings. The experimental results show that the multimodal cross-attention mechanism model proposed in this paper has higher fault diagnosis performance than other advanced models.