Scholarlink

<p>This study presents a robust vision-based monitoring approach for detecting flame combustion states in solid oxide fuel cell (SOFC) afterburners. Industrial SOFC systems face significant challenges in flame state detection due to the inherent instability of combustion processes and fluctuating gas flow rates. To address these issues and enhance monitoring reliability, we introduce FlameNet-MSF, an innovative deep learning framework that integrates multi-scale feature fusion. The architecture adopts a dual-branch design: a coarse-grained branch to extract global flame characteristics and a fine-grained branch to capture local pattern details. These complementary features are adaptively fused through a dedicated fusion module, enabling accurate flame state monitoring under complex operating conditions. Extensive experiments conducted on industrial SOFC datasets demonstrate the superior performance of FlameNet-MSF, achieving an overall classification accuracy of 99.21%, with recognition accuracies exceeding 96.9% across all three flame states. Furthermore, the framework supports real-time processing with a latency of just 29.3 ms per frame. Cross-dataset validation and ablation studies further validate the robustness and generalization capabilities of the proposed method. By providing a reliable and practical solution for automated flame monitoring, the FlameNet-MSF framework contributes to improved combustion efficiency and operational safety in industrial SOFC 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

Robust multi-scale feature fusion network for flame state monitoring in solid oxide fuel cell afterburners

  • 期刊论文

  • 发表日期：

    2025年11月5日

查看全文
  • Jingjing Wang,
  • Shuyu Zhang,
  • Jinggang Lai,
  • Jung-Sik Kim,
  • Chun Zou,
  • Zhonghua Deng,
  • Jiashu Jin,
  • Yuanwu Xu,
  • Xi Li

摘要

This study presents a robust vision-based monitoring approach for detecting flame combustion states in solid oxide fuel cell (SOFC) afterburners. Industrial SOFC systems face significant challenges in flame state detection due to the inherent instability of combustion processes and fluctuating gas flow rates. To address these issues and enhance monitoring reliability, we introduce FlameNet-MSF, an innovative deep learning framework that integrates multi-scale feature fusion. The architecture adopts a dual-branch design: a coarse-grained branch to extract global flame characteristics and a fine-grained branch to capture local pattern details. These complementary features are adaptively fused through a dedicated fusion module, enabling accurate flame state monitoring under complex operating conditions. Extensive experiments conducted on industrial SOFC datasets demonstrate the superior performance of FlameNet-MSF, achieving an overall classification accuracy of 99.21%, with recognition accuracies exceeding 96.9% across all three flame states. Furthermore, the framework supports real-time processing with a latency of just 29.3 ms per frame. Cross-dataset validation and ablation studies further validate the robustness and generalization capabilities of the proposed method. By providing a reliable and practical solution for automated flame monitoring, the FlameNet-MSF framework contributes to improved combustion efficiency and operational safety in industrial SOFC applications.