Scholarlink

<p>Accurate identification of aero-engine bearing faults through vibration signals is strongly linked to the engine system's overall performance. The Vision Transformer (ViT) is able to classify according to different images, but it is not applicable to resource-limited scenarios due to the low data utilization of ViT, which increases computation and training time. To tackle this issue, this paper proposes an application of a model based on the Data-efficient image transformer (Deit) algorithm in deep learning, which aims to improve the accuracy and efficiency of bearing fault image classification. Deit utilizes a teacher-student modeling strategy, using RegNetY-160 as the teacher model, to distill knowledge from the powerful teacher model to the smaller student model, and to guide the student model to complete the classification task, avoiding the dilemma of ViT, which requires a lot of data and computation. Meanwhile, this distillation mechanism greatly improves the training efficiency. The method's effectiveness is confirmed through experiments conducted on two public bearing datasets, where the Deit model demonstrates superior classification accuracy compared to some traditional deep learning 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

Efficient vision transformer: application of data-efficient image transformer for aero engine bearing fault classification

  • 期刊论文

  • 发表日期：

    2025年5月6日

查看全文
  • Xin Deng,
  • XuBing Fang,
  • GangJin Huang,
  • JunHeng Fu

摘要

Accurate identification of aero-engine bearing faults through vibration signals is strongly linked to the engine system's overall performance. The Vision Transformer (ViT) is able to classify according to different images, but it is not applicable to resource-limited scenarios due to the low data utilization of ViT, which increases computation and training time. To tackle this issue, this paper proposes an application of a model based on the Data-efficient image transformer (Deit) algorithm in deep learning, which aims to improve the accuracy and efficiency of bearing fault image classification. Deit utilizes a teacher-student modeling strategy, using RegNetY-160 as the teacher model, to distill knowledge from the powerful teacher model to the smaller student model, and to guide the student model to complete the classification task, avoiding the dilemma of ViT, which requires a lot of data and computation. Meanwhile, this distillation mechanism greatly improves the training efficiency. The method's effectiveness is confirmed through experiments conducted on two public bearing datasets, where the Deit model demonstrates superior classification accuracy compared to some traditional deep learning models.