Scholarlink

<p>Civil and industrial fire events have severe and irreversible repercussions, which include economic damage as well as loss of life. Therefore, a fire detection system (FDS) is needed that should have high reliability along with a prompt response. Traditional fire detection models, which mostly look for accuracy, can be misled by imbalanced datasets, which leads to overlooking the consequences of misclassification. This could be critical during a fire outbreak in infrastructure such as civil and industrial facilities. The proposed study mainly focuses on the cost-sensitive learning (CSL) approach to increase the reliability and quick response of the FDS. This has been achieved by reducing the misclassification costs related to false positives (FPs) and false negatives (FNs). This study mainly examines popular deep learning models such as YOLO, RCNN, and SSD to evaluate their impact on FPs and FNs in FDS. Experimental results demonstrate that the achieved precision and recall of 0.98 while maintaining an inference time of 24&#xa0;ms. Compared with baseline YOLOv7, FPs were reduced from 41 to 15 and FNs from 31 to 13 which substantially lower the misclassification cost. These findings confirm that CSL enhances both accuracy and reliability in real-time, high-stakes scenarios. This research highlights the importance of integrating cost factors into the training process, leading to more effective models for critical fire detection applications.</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

Real-Time Fire Detection System for Critical Infrastructure with Cost-Sensitive Learning

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Ram Pravesh,
  • Bikash Chandra Sahana

摘要

Civil and industrial fire events have severe and irreversible repercussions, which include economic damage as well as loss of life. Therefore, a fire detection system (FDS) is needed that should have high reliability along with a prompt response. Traditional fire detection models, which mostly look for accuracy, can be misled by imbalanced datasets, which leads to overlooking the consequences of misclassification. This could be critical during a fire outbreak in infrastructure such as civil and industrial facilities. The proposed study mainly focuses on the cost-sensitive learning (CSL) approach to increase the reliability and quick response of the FDS. This has been achieved by reducing the misclassification costs related to false positives (FPs) and false negatives (FNs). This study mainly examines popular deep learning models such as YOLO, RCNN, and SSD to evaluate their impact on FPs and FNs in FDS. Experimental results demonstrate that the achieved precision and recall of 0.98 while maintaining an inference time of 24 ms. Compared with baseline YOLOv7, FPs were reduced from 41 to 15 and FNs from 31 to 13 which substantially lower the misclassification cost. These findings confirm that CSL enhances both accuracy and reliability in real-time, high-stakes scenarios. This research highlights the importance of integrating cost factors into the training process, leading to more effective models for critical fire detection applications.