Scholarlink

<p>The motion anomalies of needle-selection blades in electromagnetic needle selectors are subtle and rare. Due to limited observability, difficulty in feature extraction, and severe class imbalance caused by scarce fault samples, data-driven models struggle with training and generalization. To address this, a physics-informed fault diagnosis method for electromagnetic needle selectors is proposed for rare fault detection using visually captured trajectory data, combining mechanism-guided trajectory synthesis with a fault–feature relevance-weighted random forest (RW-RF) classifier. High-fidelity minority-class fault trajectories are generated using a calibrated dynamic model specifically developed for the electromagnetic needle selector, whose key parameters are adjusted according to observed physical failure mechanisms to mitigate class imbalance. The RW-RF defines fault–feature correlation for each fault class as the inverse of intra-class dispersion, and uses these class-specific relevance scores to modulate entropy-based information gain in the splitting criterion, thereby prioritizing discriminative features during tree construction. Built on physics-informed data augmentation, the method achieves WAP, WAR, and WAF scores of 97%, with an F1-score of 0.84 and recall of 92% and 80% for the two rare fault classes, outperforming mainstream data augmentation and classification methods and enabling efficient rare-fault detection under small-sample conditions.</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

Vision to detection: physics-informed data augmentation and relevance-weighted random forests for electromagnetic needle selection anomaly detection

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Laihu Peng,
  • Song Liu,
  • Yubao Qi,
  • Xin Ru,
  • Kaiyuan Shao

摘要

The motion anomalies of needle-selection blades in electromagnetic needle selectors are subtle and rare. Due to limited observability, difficulty in feature extraction, and severe class imbalance caused by scarce fault samples, data-driven models struggle with training and generalization. To address this, a physics-informed fault diagnosis method for electromagnetic needle selectors is proposed for rare fault detection using visually captured trajectory data, combining mechanism-guided trajectory synthesis with a fault–feature relevance-weighted random forest (RW-RF) classifier. High-fidelity minority-class fault trajectories are generated using a calibrated dynamic model specifically developed for the electromagnetic needle selector, whose key parameters are adjusted according to observed physical failure mechanisms to mitigate class imbalance. The RW-RF defines fault–feature correlation for each fault class as the inverse of intra-class dispersion, and uses these class-specific relevance scores to modulate entropy-based information gain in the splitting criterion, thereby prioritizing discriminative features during tree construction. Built on physics-informed data augmentation, the method achieves WAP, WAR, and WAF scores of 97%, with an F1-score of 0.84 and recall of 92% and 80% for the two rare fault classes, outperforming mainstream data augmentation and classification methods and enabling efficient rare-fault detection under small-sample conditions.