Scholarlink

<p>Ensuring the operational reliability of high-voltage insulators is fundamental to power system security. However, conventional supervised detection methods are constrained by the high cost of pixel-level annotation and the inherent scarcity of fault samples in industrial scenarios. To address these issues, this study proposes a Context-Aware Multi-scale Memory-augmented Autoencoder (CAM-MAE) for high-precision visual anomaly mining without exposure to fault data. Distinct from existing memory-based approaches that suffer from background interference, the proposed framework incorporates a multi-scale feature pyramid to capture cross-scale defect patterns and utilizes a query-specific memory bank mechanism. By constraining the latent space to a linear combination of learned normal prototypes, the model forces anomalous inputs to yield pronounced reconstruction residuals. Additionally, a novel context-aware gating module is integrated to suppress background noise through spatial semantic constraints. Empirical evaluations on the Chinese Power Line Insulator Dataset and an external cross-domain dataset demonstrate that CAM-MAE achieves an AUROC of 98.4% and an F1-score of 0.96 in an unsupervised setting, significantly outperforming state-of-the-art methods including GAN-based and Transformer-based models. These results establish CAM-MAE as a reliable, interpretable, and computationally efficient framework for next-generation intelligent transmission line 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

Unsupervised visual anomaly detection for high voltage insulators using context aware multi scale memory networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月5日

查看全文
  • Zongxi Xie,
  • Song Wu,
  • Ting Yang,
  • Na Zhan,
  • Yukang Li

摘要

Ensuring the operational reliability of high-voltage insulators is fundamental to power system security. However, conventional supervised detection methods are constrained by the high cost of pixel-level annotation and the inherent scarcity of fault samples in industrial scenarios. To address these issues, this study proposes a Context-Aware Multi-scale Memory-augmented Autoencoder (CAM-MAE) for high-precision visual anomaly mining without exposure to fault data. Distinct from existing memory-based approaches that suffer from background interference, the proposed framework incorporates a multi-scale feature pyramid to capture cross-scale defect patterns and utilizes a query-specific memory bank mechanism. By constraining the latent space to a linear combination of learned normal prototypes, the model forces anomalous inputs to yield pronounced reconstruction residuals. Additionally, a novel context-aware gating module is integrated to suppress background noise through spatial semantic constraints. Empirical evaluations on the Chinese Power Line Insulator Dataset and an external cross-domain dataset demonstrate that CAM-MAE achieves an AUROC of 98.4% and an F1-score of 0.96 in an unsupervised setting, significantly outperforming state-of-the-art methods including GAN-based and Transformer-based models. These results establish CAM-MAE as a reliable, interpretable, and computationally efficient framework for next-generation intelligent transmission line inspection.