Scholarlink

<p>Gearboxes are critical mechanical components in advanced manufacturing systems, rotating machinery, and industrial production lines where unplanned failures lead to costly downtime and disrupted manufacturing processes. While data augmentation increases dataset size, it does not increase the information density of the dataset and can lead to controlled overfitting by exposing the model to artificial variability rather than genuine fault representation. This study investigates a data-centric alternative based on data loading strategies, focusing on a special extended case of selective embedding that reorganizes synchronized multi-axis vibration signals without modifying their physical content. Under strict leakage safe file level k-fold cross validation, single channel, multi channel, and the proposed selective embedding strategy were evaluated across 5, 10, 15, and 20 class gearbox classification tasks using identical segmentation, FFT transformation, and ResNet-8 training settings. Results show that selective embedding achieves the highest accuracy and improved stability for small label classification (5 and 10 classes) while maintaining training time comparable to single channel loading. As task granularity increases to 15 and 20 classes, multi channel loading becomes more advantageous due to the importance of joint cross-axis information. The findings demonstrate that data presentation directly influences learning performance and should be considered alongside model architecture and signal processing in vibration gearbox condition monitoring.</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

Influence of data loading strategies on multi sensor vibration gearbox fault diagnosis under limited data conditions

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Mert Sehri,
  • Tongtong Yan,
  • Govind Vashishtha,
  • Sumika Chauhan

摘要

Gearboxes are critical mechanical components in advanced manufacturing systems, rotating machinery, and industrial production lines where unplanned failures lead to costly downtime and disrupted manufacturing processes. While data augmentation increases dataset size, it does not increase the information density of the dataset and can lead to controlled overfitting by exposing the model to artificial variability rather than genuine fault representation. This study investigates a data-centric alternative based on data loading strategies, focusing on a special extended case of selective embedding that reorganizes synchronized multi-axis vibration signals without modifying their physical content. Under strict leakage safe file level k-fold cross validation, single channel, multi channel, and the proposed selective embedding strategy were evaluated across 5, 10, 15, and 20 class gearbox classification tasks using identical segmentation, FFT transformation, and ResNet-8 training settings. Results show that selective embedding achieves the highest accuracy and improved stability for small label classification (5 and 10 classes) while maintaining training time comparable to single channel loading. As task granularity increases to 15 and 20 classes, multi channel loading becomes more advantageous due to the importance of joint cross-axis information. The findings demonstrate that data presentation directly influences learning performance and should be considered alongside model architecture and signal processing in vibration gearbox condition monitoring.