Scholarlink

<p>This paper presents C2DAN, a cross-domain anomaly detection strategy for distributed fiber optic sensing (DFOS) systems. Due to the long deployment range of optical fibers, variations in installation environments often cause the same type of event to exhibit different signal patterns across locations, resulting in significant domain shifts. Additionally, subtle differences in fiber-medium responses to identical external forces may lead to overlapping feature representations between different classes, thereby inducing class confusion. To address these challenges, C2DAN builds upon a CNN-Transformer backbone and incorporates an entropy-weighted conditional adversarial mechanism. Specifically, the extracted feature representation and the corresponding class prediction distribution are combined through a multilinear mapping to construct a joint feature-prediction representation, which is then fed into the domain discriminator. This design enables class-conditional domain alignment rather than marginal distribution alignment alone. In addition, entropy-based weighting is introduced to reduce the influence of uncertain samples during adversarial training, thereby improving the stability and semantic reliability of cross-domain alignment. Furthermore, a SupCon Loss is introduced to reduce intra-class variance and enlarge inter-class margins in feature space, enhancing the model’s ability to distinguish between overlapping classes caused by signal response variations. Extensive experiments under different deployment scenarios and operational conditions demonstrate that C2DAN outperforms existing methods in both accuracy and robustness, maintaining stable anomaly detection performance under domain shifts and thereby enhancing the reliability of DFOS-based monitoring systems.</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

C2DAN: conditional domain adversarial network with supcon loss for mitigating class confusion in anomaly detection of DFOS

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Ning Xu,
  • Yiwei Li,
  • Fuyang Chen,
  • Ruihang Xu

摘要

This paper presents C2DAN, a cross-domain anomaly detection strategy for distributed fiber optic sensing (DFOS) systems. Due to the long deployment range of optical fibers, variations in installation environments often cause the same type of event to exhibit different signal patterns across locations, resulting in significant domain shifts. Additionally, subtle differences in fiber-medium responses to identical external forces may lead to overlapping feature representations between different classes, thereby inducing class confusion. To address these challenges, C2DAN builds upon a CNN-Transformer backbone and incorporates an entropy-weighted conditional adversarial mechanism. Specifically, the extracted feature representation and the corresponding class prediction distribution are combined through a multilinear mapping to construct a joint feature-prediction representation, which is then fed into the domain discriminator. This design enables class-conditional domain alignment rather than marginal distribution alignment alone. In addition, entropy-based weighting is introduced to reduce the influence of uncertain samples during adversarial training, thereby improving the stability and semantic reliability of cross-domain alignment. Furthermore, a SupCon Loss is introduced to reduce intra-class variance and enlarge inter-class margins in feature space, enhancing the model’s ability to distinguish between overlapping classes caused by signal response variations. Extensive experiments under different deployment scenarios and operational conditions demonstrate that C2DAN outperforms existing methods in both accuracy and robustness, maintaining stable anomaly detection performance under domain shifts and thereby enhancing the reliability of DFOS-based monitoring systems.