Scholarlink

<p>Scarce or unseen data makes industrial equipment databases unable to cover all potential states and faults, restricting deep learning in intelligent maintenance. This paper thus proposes a zero-shot state recognition method integrating few-shot augmentation, self-attention, and adaptive thresholds. First, a threshold-assigned self-attention convolutional neural network (TA-SACNN) with a self-attention mechanism and threshold adaptive adjustment is pre-trained on seen classes to derive thresholds. Second, unseen classes are used as input and judged according to the thresholds, with new classes used as output and subjected to secondary filtering. In addition, a novel autoencoder loss function has been proposed for the augmentation of new class data. Finally, the new class data and augmented data are treated as seen classes. Whenever new class data is collected, the above process is repeated, expanding the database and establishing a fault diagnosis model. This method was validated on the HUST bearing dataset and the PHM2009 gearbox dataset, demonstrating scalability and high recognition accuracy, making it suitable for online intelligent maintenance of industrial equipment.</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 novel industrial equipment fault diagnosis method combining unseen class recognition and few-shot augmentation

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Qianhui Chen,
  • Zhongjun Yin,
  • Tian Han,
  • Jiachen Pang

摘要

Scarce or unseen data makes industrial equipment databases unable to cover all potential states and faults, restricting deep learning in intelligent maintenance. This paper thus proposes a zero-shot state recognition method integrating few-shot augmentation, self-attention, and adaptive thresholds. First, a threshold-assigned self-attention convolutional neural network (TA-SACNN) with a self-attention mechanism and threshold adaptive adjustment is pre-trained on seen classes to derive thresholds. Second, unseen classes are used as input and judged according to the thresholds, with new classes used as output and subjected to secondary filtering. In addition, a novel autoencoder loss function has been proposed for the augmentation of new class data. Finally, the new class data and augmented data are treated as seen classes. Whenever new class data is collected, the above process is repeated, expanding the database and establishing a fault diagnosis model. This method was validated on the HUST bearing dataset and the PHM2009 gearbox dataset, demonstrating scalability and high recognition accuracy, making it suitable for online intelligent maintenance of industrial equipment.