Scholarlink

<p>Railway fasteners are critical components of railway infrastructure, ensuring the safety and reliability of trains. However, detecting defects in these fasteners is challenging due to the rarity of defective samples. To address this, we introduce a self-supervised contrastive learning approach that extracts meaningful representations from unlabeled data, minimizing dependence on large labeled datasets. Our method integrates deep metric learning with distance measures such as Euclidean, Minkowski, and Cosine Similarity to enhance anomaly detection. Unlike traditional supervised models, which often fail to generalize with limited defective data, our approach effectively distinguishes anomalies without direct training on defective samples. Additionally, we employ advanced data augmentation techniques and construct a real-world point cloud dataset of railway fasteners, improving model robustness and generalization. Experimental results demonstrate that our method achieves improvement in anomaly detection and reduction in false positives compared to existing techniques, establishing a scalable and reliable solution for railway fastener inspection. </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

Self-supervised contrastive anomaly detection in railway fasteners using point clouds and deep metric learning for imbalance dataset

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Qasim Zaheer,
  • Jin Wang,
  • S. Muhammad Ahmed Hassan Shah,
  • Haleema Ehsan,
  • Syed Faizan Hussain Shah,
  • Chengbo Ai,
  • Jiakai Kuang,
  • Weidong Wang,
  • Shi Qiu

摘要

Railway fasteners are critical components of railway infrastructure, ensuring the safety and reliability of trains. However, detecting defects in these fasteners is challenging due to the rarity of defective samples. To address this, we introduce a self-supervised contrastive learning approach that extracts meaningful representations from unlabeled data, minimizing dependence on large labeled datasets. Our method integrates deep metric learning with distance measures such as Euclidean, Minkowski, and Cosine Similarity to enhance anomaly detection. Unlike traditional supervised models, which often fail to generalize with limited defective data, our approach effectively distinguishes anomalies without direct training on defective samples. Additionally, we employ advanced data augmentation techniques and construct a real-world point cloud dataset of railway fasteners, improving model robustness and generalization. Experimental results demonstrate that our method achieves improvement in anomaly detection and reduction in false positives compared to existing techniques, establishing a scalable and reliable solution for railway fastener inspection.