Real-time monitoring of coke particle size in blast furnace tuyere via an improved object detection algorithm

Abstract

In the blast furnace (BF) production process, the particle size variation trend of tuyere coke can be utilised to optimise operations and stabilise furnace conditions. However, due to the high-temperature and high-pressure characteristics of the BF itself, real-time and effective sampling and monitoring of coke particles remain challenging. This study innovatively introduces machine vision and deep learning technologies to achieve online detection of tuyere coke particle size distribution. To address issues such as limited detection methods for BF tuyere coke, low accuracy in small-to-medium target detection, high missed detection rates, and poor real-time performance, this study proposes an improved CTD-YOLO (Tuyere Coke Target Detection YOLO) algorithm based on YOLOv5. Four key improvements were implemented over YOLOv5: 1) Integration of FasterNet architecture to reduce redundant computations and memory access for efficient spatial feature extraction. 2) Incorporation of Squeeze-and-Excitation Network attention mechanism for feature-wise adaptive weighting. 3) Adoption of Attention-based Intra-scale Feature Interaction self-attention mechanism for multi-scale feature fusion. 4) Parameter balancing strategy to optimise computational resource allocation. The optimised CTD-YOLO network achieved 2.2% improvement in both mean Average Precision (mAP) mAP@0.5 and mAP@0.5:0.95 metrics, 1.5% increase in F1-score, and 46.88% reduction in GFLOPs. The final particle size recognition results demonstrate strong alignment with physical sampling data from the raceway zone, establishing a novel machine vision-based method for online granularity detection in complex industrial environments. Considering the stability requirements for industrial applications, a multi-stage validation process was conducted to comprehensively evaluate the model. Through iterative parameter optimisation, an effective balance was achieved among computational cost, accuracy, and stability. The model is now ready for real-time online detection in industrial field applications.