A New Monitoring Method for the Injection Volume of Blast Furnace Clay Gun Based on Object Detection

Xunkai Zhang, Helan Liang, Hongwei Guo, Bingji Yan, Hao XuProcesses 2025 · DOI

TL;DR

This paper introduces a novel computer vision approach for accurately monitoring the injection volume of clay guns in blast furnaces, a critical but traditionally manual process. They developed a new object detection model, PSCfrcn, incorporating multi-stage positional encoding and a two-branch self-challenge mechanism, to overcome the challenges posed by harsh industrial environments and improve reliability.

What Problem Does This Paper Address?

Blast furnace ironmaking relies on clay guns to seal tapholes after hot metal discharge. Accurately measuring the volume of clay injected is crucial: too much can delay production, while too little can damage the furnace. The current method involves manual observation of a vernier scale, which is prone to errors due to poor visibility (smoke, dust), inconsistent operator skill, and high delay. This manual approach leads to unreliable data and potential operational inefficiencies. The paper aims to automate this monitoring process using computer vision to provide a more reliable and timely solution.

Methods

The core of the proposed solution is an improved object detection model called PSCfrcn, built upon the Faster R-CNN architecture. Key innovations include:

  • Interpolation Annotation Technique: To address the difficulty of precisely labeling the vernier's position across a continuous range, the authors developed a method to interpolate labels between manually annotated integer and half-scale points. This creates a more comprehensive dataset for training.
  • Multi-Stage Positional Encoding (PE): Standard positional encoding helps models understand spatial relationships. This paper enhances it by applying PE across multiple stages of the ResNet backbone (specifically, the last three stages). This allows the model to better capture the relative positions of local features at different scales, which is crucial for inferring the vernier's position relative to the scales.
  • Two-Branch Self-Challenge (SC): To combat the adverse environmental conditions (shaking, poor lighting, dust, water vapor), a self-challenge module is introduced. This module adaptively suppresses certain features during training, forcing the model to be more robust and generalize better. It's implemented with two branches: one for classification and one for regression, allowing for tailored feature selection for each task.

Datasets

The authors created a new dataset specifically for this task.

  • Clay Gun Dataset: This dataset was collected from videos of a blast furnace clay gun operation at Tranvic Steel Co., Ltd. It captures various vernier positions under different working conditions. The annotation process involved manual labeling of integer and half-scale positions, followed by linear interpolation to generate labels for intermediate scales. The dataset was split into training and validation sets in a 4:1 ratio at the video level to prevent data leakage.

Key Insights and Ideas

  • Computer Vision for Industrial Monitoring: The paper demonstrates the viability of applying advanced computer vision techniques, specifically object detection, to solve a critical industrial monitoring problem that has traditionally relied on manual methods.
  • Environmental Robustness is Key: The harsh conditions in a blast furnace environment (dust, steam, vibration) are significant challenges for computer vision. The proposed self-challenge mechanism is crucial for achieving reliable performance in such settings.
  • Multi-Scale Feature Importance: Capturing the vernier's position relative to the scales requires understanding features at different levels of abstraction. The multi-stage positional encoding effectively addresses this by integrating positional information across multiple ResNet stages.
  • Two-Stage Detectors Shine Here: The authors found that two-stage detectors like Faster R-CNN are better suited for this task than one-stage detectors (like YOLO) due to the need for high accuracy in complex, noisy environments, even though one-stage detectors are generally faster.

Key Results

  • The proposed PSCfrcn model, particularly the version with a ResNet34 backbone (PSCfrcn-res34), achieved the highest accuracy (90.58%) and mAP (84.53%) among all tested models on their custom dataset.
  • This represents a significant improvement of 3.35% in accuracy and 2.93% in mAP compared to the best baseline (Faster R-CNN + with ResNet50).
  • The model's inference time of 0.128 seconds (for PSCfrcn-Res34) is considered sufficient for real-time monitoring, meeting the practical requirement of processing data at approximately one second intervals.
  • The 1st deriv. SD metric, used to evaluate the smoothness of the vernier's predicted trajectory, also showed favorable results for the proposed model, indicating less jitter.

Limitations and Future Work

  • Annotation Challenges: The authors acknowledge that manual annotation for intermediate scale categories can be challenging and may introduce some errors, potentially impacting the model's performance on these specific categories.
  • Data Imbalance: The dataset exhibits an imbalance in the amount of training data for different scale ranges, with less data for scales below 1.5 and above 4. This leads to a decline in performance for these extreme ranges.
  • Future Directions: The authors suggest several avenues for future research:
  • Exploring fine-grained classification methods to further improve detection accuracy.
  • Developing algorithms to filter out obviously erroneous results.
  • Implementing techniques like moving averages to enhance the stability of the predictions.

Summary updated

Abstract

Monitoring the injection volume of clay guns is important in blast furnace ironmaking. Currently, such monitoring data are often recorded manually, which has limitations such as low reliability and high delay. To address these issues, we revisit the task from a computer vision perspective and propose an object detection method. First, we introduce an interpolation annotation technique to build a clay gun dataset. With it as a foundation, an improved object detection model called Faster R-CNN with multi-stage Positional encoding and two-branch Self Challenge (PSCfrcn) is proposed. Our model leverages multi-stage positional encoding (PE) to focus more on the relative position of local features, and a self-challenge (SC) module to mitigate the interference caused by the harsh environments. We conduct extensive experiments on the clay gun dataset from actual industrial scenarios, and use various metrics to validate the performance. Experimental results demonstrate that our method can significantly enhance the model’s discriminative power and generalization ability, which is a promising direction for this task.
BibTeX
@inproceedings{zhang2025new,
  author = {Xunkai Zhang and Helan Liang and Hongwei Guo and Bingji Yan and Hao Xu},
  title = {{A New Monitoring Method for the Injection Volume of Blast Furnace Clay Gun Based on Object Detection}},
  booktitle = {Processes},
  year = {2025},
  doi = {10.3390/pr13041006},
}
Figure