Scholarlink

<p>Substation Equipment Infrared Image (SEII) object detection is crucial for real-time fault diagnosis in the Industrial Internet of Things (IoT). However, deploying deep learning models on resource-constrained embedded devices faces significant challenges, primarily due to the low contrast of infrared imagery and the high computational cost of modern detectors, which often leads to suboptimal accuracy and high latency. To address these issues, this study presents SEII-YOLO, a lightweight architecture optimized based on the YOLOv11n framework. First, to enhance feature extraction capabilities in complex thermal backgrounds, a Convolutional Block Attention Module (CBAM) is integrated into the backbone. Second, to balance efficiency and performance, the computationally intensive C2PSA layer is removed, reducing model complexity. Experimental results demonstrate that SEII-YOLO achieves a precision of 93.5%, a recall of 82.0%, and an mAP50 of 88.3%. Compared to the baseline, the proposed model achieves notable performance gains—particularly in distinguishing similar-looking equipment—while reducing parameters by 8.96% and computational load by 3.17%. Furthermore, the model deployed on an RK3588S embedded platform utilizing NPU acceleration achieves an inference speed of 66.8 FPS. These results confirm the model’s strong viability for practical real-time edge computing applications. To facilitate reproducibility and further research, the code has been made publicly available at <a href="https://github.com/wityou/SEII-YOLO">https://github.com/wityou/SEII-YOLO</a>.</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

SEII-YOLO: a lightweight enhanced object detection model for substation equipment infrared images with applications in IoT embedded devices

  • 期刊论文

  • 发表日期：

    2026年7月15日

查看全文
  • Guihong Liu,
  • Xiaoning Jiang,
  • Mengbo Ju,
  • Sudan Liao,
  • Zhigang Gan,
  • Binxiao Yu

摘要

Substation Equipment Infrared Image (SEII) object detection is crucial for real-time fault diagnosis in the Industrial Internet of Things (IoT). However, deploying deep learning models on resource-constrained embedded devices faces significant challenges, primarily due to the low contrast of infrared imagery and the high computational cost of modern detectors, which often leads to suboptimal accuracy and high latency. To address these issues, this study presents SEII-YOLO, a lightweight architecture optimized based on the YOLOv11n framework. First, to enhance feature extraction capabilities in complex thermal backgrounds, a Convolutional Block Attention Module (CBAM) is integrated into the backbone. Second, to balance efficiency and performance, the computationally intensive C2PSA layer is removed, reducing model complexity. Experimental results demonstrate that SEII-YOLO achieves a precision of 93.5%, a recall of 82.0%, and an mAP50 of 88.3%. Compared to the baseline, the proposed model achieves notable performance gains—particularly in distinguishing similar-looking equipment—while reducing parameters by 8.96% and computational load by 3.17%. Furthermore, the model deployed on an RK3588S embedded platform utilizing NPU acceleration achieves an inference speed of 66.8 FPS. These results confirm the model’s strong viability for practical real-time edge computing applications. To facilitate reproducibility and further research, the code has been made publicly available at https://github.com/wityou/SEII-YOLO.