对恶劣环境下钢包运行动作姿态信息的有效感知是钢铁安全生产管控智能化需要解决的重要问题。总结钢包运行的时序信息特征，将钢包运行时的复杂安全信息分解为一系列可识别的基础动作，在此基础上构建钢包动作识别数据集。选用时序动作检测模型识别钢包动作信息，并根据钢包运行的视觉特性及精准定位需求将原网络目标检测分支替换为改进后的你只看一次(You Only Look Once，YOLOv8)图像分割模型。试验结果表明，改进后的模型所占存储容量减少63。98％，计算需求降低40。6％;识别准确率和召回率分别提高了0。95％与0。51％，且mAP50达到98。6％，能满足钢包实时精准定位的需求。改进后的时序动作检测模型各类动作平均识别准确率达到87。44％。研究表明，所提出时空动作检测改进模型能有效检测复杂环境内钢包的位置信息与基础动作信息，可以满足钢包复杂工序识别、目标追踪、碰撞预警、倾覆洞穿预警等安全检测任务的需求，降低安全管控所需的人力物力成本。
Deep learning-based sensing of ladle motion and location information
Effective perception of the movement and posture information of steel ladles in harsh operational environments is a crucial technical challenge that needs to be addressed to advance the intelligent control of steel production.This paper conducts an in-depth study of the temporal information characteristics of steel ladle operations,breaking down complex safety information into a series of recognizable basic actions,thereby constructing a comprehensive dataset for ladle action recognition.Unlike simple motion detection,this research emphasizes the necessity of spatiotemporal action detection to ensure the dynamism and continuity of ladle movements,which is vital for making precise real-time responses in the safety-critical environment of steel manufacturing.Utilizing the unique visual features and precise localization requirements of steel ladle operations,the paper identifies ladle action and posture information.The spatiotemporal action detection model employs an improved YOLOv8 image segmentation model,replacing the original network's target detection branch,thereby enhancing the model's ability to effectively process spatial and temporal data.Experimental results show significant improvements in the model:a 63.98％reduction in size,a 40.6％decrease in computational demand,a 0.95％increase in recognition accuracy,a 0.51％increase in recall rates,and a mAP50 value of 98.6％.These enhancements meet the stringent requirements for real-time and precise localization of the ladle.Additionally,the improved temporal action detection model achieves an average action recognition accuracy of 87.44％.Comparative experiments demonstrate the model's capability to accurately detect the position and basic action information of ladles in complex environments.This capability is crucial for managing complex process identification,target tracking,collision warnings,and warnings for potential overturning and piercing accidents,thereby optimizing safety detection tasks and significantly reducing the human and material resources required for safety control.The deep learning techniques and methodologies developed in this study explore a framework for ladle posture and action recognition,facilitating the intelligent and digital transformation of safety production practices in the steel industry.This approach underscores the importance of integrating minimal temporal action elements into behavior detection,which is crucial for proactive safety management and the intelligentization of safety in real-time operational environments.
safety engineeringsafety in the metallurgical industryimage segmentationaction recognitionladle action detectionsafety intelligence
万方数据
维普
