Scholarlink

<p>This paper presents an innovative Transformer-based deep learning strategy for optimizing the placement of sensors aiming at structural health monitoring of critical mechanical failures in semiconductor probe cards. Failures in probe cards, including substrate cracks and loosened screws, critically affect semiconductor manufacturing yield and reliability. Sensors placement on the probe card enables early failure detection and preventive maintenance. However, traditional monitoring approaches are limited by the severe space constraints of the probe head, which prevent the installation of dense sensor arrays. To overcome this, we employ a Digital Shadow methodology to optimize sensor placement. Utilizing Finite Element simulations we generated a comprehensive dataset of vibration responses under various health states, specifically 125 physics-informed scenario variants encompassing material property variations, diverse environmental thermal states, and varying mechanical loading conditions. We propose TransformerSHM, a hybrid architecture that leverages multi-head attention mechanisms to not only classify failure modes but also automatically identify the most informative sensor locations. The methodology comprises: (1) physics-aware data augmentation with validated noise levels matching sensor specifications; (2) rigorous <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(3\times 10\)</EquationSource> </InlineEquation>-fold stratified cross-validation with 30 model runs; and (3) hyperparameter optimization via grid search identifying the optimal configuration. The model achieves 98.75% accuracy in classifying health states on the simulated dataset, with exceptional 98.7% crack detection recall. Crucially, the attention weights provide a data-driven ranking of sensor importance, validated through 100% overlap with classical Fisher Information Matrix methods, allowing for the reduction of the monitoring network from 28 candidate locations to a minimal optimal subset of 5–8 ones. These results demonstrate the methodological potential of attention-based deep learning for sensor placement optimization of probe cards within physics-based simulations, highlighting pathways for future high-precision manufacturing.</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

Transformer-based approach to optimal sensor placement for structural health monitoring of probe cards

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月30日

查看全文
  • Mehdi Bejani,
  • Marco Mauri,
  • Simone Todaro,
  • Stefano Mariani

摘要

This paper presents an innovative Transformer-based deep learning strategy for optimizing the placement of sensors aiming at structural health monitoring of critical mechanical failures in semiconductor probe cards. Failures in probe cards, including substrate cracks and loosened screws, critically affect semiconductor manufacturing yield and reliability. Sensors placement on the probe card enables early failure detection and preventive maintenance. However, traditional monitoring approaches are limited by the severe space constraints of the probe head, which prevent the installation of dense sensor arrays. To overcome this, we employ a Digital Shadow methodology to optimize sensor placement. Utilizing Finite Element simulations we generated a comprehensive dataset of vibration responses under various health states, specifically 125 physics-informed scenario variants encompassing material property variations, diverse environmental thermal states, and varying mechanical loading conditions. We propose TransformerSHM, a hybrid architecture that leverages multi-head attention mechanisms to not only classify failure modes but also automatically identify the most informative sensor locations. The methodology comprises: (1) physics-aware data augmentation with validated noise levels matching sensor specifications; (2) rigorous \(3\times 10\) -fold stratified cross-validation with 30 model runs; and (3) hyperparameter optimization via grid search identifying the optimal configuration. The model achieves 98.75% accuracy in classifying health states on the simulated dataset, with exceptional 98.7% crack detection recall. Crucially, the attention weights provide a data-driven ranking of sensor importance, validated through 100% overlap with classical Fisher Information Matrix methods, allowing for the reduction of the monitoring network from 28 candidate locations to a minimal optimal subset of 5–8 ones. These results demonstrate the methodological potential of attention-based deep learning for sensor placement optimization of probe cards within physics-based simulations, highlighting pathways for future high-precision manufacturing.