Scholarlink

<p>Single-algorithm visual defect detection in substations suffers from high false positives and missed detections under varying illumination, occlusion, and imbalanced samples. To address this, we propose a multi-algorithm cross-validation framework that fuses YOLOv8, Mask R-CNN, and Swin Transformer in parallel to leverage their complementary inductive biases. The framework integrates a strict spatial alignment via IoU-based matching, a consistency arbitration layer based on majority voting and confidence-driven selection, and a dynamic confidence-weighted fusion scheme to suppress individual model noise and systematic bias. On a public substation defect dataset, the method achieves mAP@0.5 ≥ 0.91 and Recall@0.5 ≥ 0.92 across all illumination conditions. Under four types of severe interference (rain/fog, reflection, dense occlusion, and electromagnetic noise), it attains a false positive rate ≤ 0.03 and a miss detection rate ≤ 0.04, outperforming single-model baselines and representative detectors. The proposed approach provides a reliable decision‑ready perception solution for unmanned substation inspection by enabling consistent, well‑calibrated outputs that directly support maintenance scheduling and risk mitigation.</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

Multi-algorithm cross-validation-driven reliability improvement method for substation defect detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Hai Xue,
  • Ziquan Liu,
  • Zhen Wang,
  • Yongling Lu,
  • Ze Yin

摘要

Single-algorithm visual defect detection in substations suffers from high false positives and missed detections under varying illumination, occlusion, and imbalanced samples. To address this, we propose a multi-algorithm cross-validation framework that fuses YOLOv8, Mask R-CNN, and Swin Transformer in parallel to leverage their complementary inductive biases. The framework integrates a strict spatial alignment via IoU-based matching, a consistency arbitration layer based on majority voting and confidence-driven selection, and a dynamic confidence-weighted fusion scheme to suppress individual model noise and systematic bias. On a public substation defect dataset, the method achieves mAP@0.5 ≥ 0.91 and Recall@0.5 ≥ 0.92 across all illumination conditions. Under four types of severe interference (rain/fog, reflection, dense occlusion, and electromagnetic noise), it attains a false positive rate ≤ 0.03 and a miss detection rate ≤ 0.04, outperforming single-model baselines and representative detectors. The proposed approach provides a reliable decision‑ready perception solution for unmanned substation inspection by enabling consistent, well‑calibrated outputs that directly support maintenance scheduling and risk mitigation.