基于计算机视觉的连铸漏钢空间-时序特征可视化及其智能预测集成方法

Breakout is a disastrous accident during continuous casting. The accuracy and speediness of breakout prediction are important to the smooth and steady continuous casting production. At present, because of the frequent false and missed alarms, the prediction method of one-dimensional temperature cannot satisfy the needs of quick and sustainable development of continuous casting. Therefore this project focuses on breakout space-time characteristics visualization and integrated prediction method during continuous casting. Firstly, a mould thermograph is developed by virtue of frame difference algorithm based on careful investigation on the influence factors of breakouts, such as mould powder and molten steel composition. Then visual space-time characteristics are extracted quickly and accurately by computer vision algorithms, such as region segmentation, run labeling. Secondly, the common visual characteristics are epurated based on the statistics of the collected breakout samples. The space propagation and dynamic evolution rule of breakouts will be revealed. Finely, a decision tree model will be adopted to prejudge the abnormal regions in order to remove some false sticker regions. Then random forest and support vector machine models will be built up, trained and tested in order to establish an accurate and efficient prediction method. The results are helpful to obtain new progress on prevention and prediction of breakout. An integrated breakout prediction method is developed by computer vision, space-time characteristics, artificial intelligence. It also improves the visualization and intelligence of abnormal detection during continuous casting.

漏钢是连铸生产的灾难性事故，及时、准确预报漏钢是稳定和保障连铸顺行的重中之重。目前频繁误报、漏报等问题普遍存在，一维温度曲线预报方法已无法满足持续、高效的连铸发展需求。因此，本项目重点围绕连铸漏钢空间-时序特征可视化和预测集成方法进行研究。首先，在细致考察保护渣、钢水成分等因素对漏钢影响和作用的基础上，借助帧间差分算法，实现结晶器温度热成像，运用区域分割、游程标记等计算机视觉算法，快速、准确提取漏钢空间-时序可视化特征。其次，基于积累的漏钢样本，统计归纳漏钢的共性可视化特征，揭示漏钢空间传播和动态演化规律。最后，采用决策树对异常区域进行预判断，剔除部分伪黏结区域，建立随机森林、支持向量机漏钢预测模型，通过训练和测试，确立准确、高效的漏钢预测方法。研究结果有望在预防、预测漏钢事故上取得新进展，开发基于视觉-空间/时序-人工智能的漏钢预测集成新方法，提升连铸过程异常检测的可视化、智能化水平。

漏钢是连铸板坯生产的重大安全事故，频繁的误报和时有的漏报严重影响了连铸生产顺行。本项目基于连铸结晶器在线监控系统和现场实验研究，从钢种、铸坯尺寸、保护渣、拉速、液位等方面，着重分析了黏结漏钢的主要影响因素，在此基础之上，利用计算机视觉检测技术，采用游程标记、区域分割等图像处理算法，对黏结漏钢异常区域进行了快速标记，并提取热区域和冷区域特征，之后，基于现场的黏结漏钢实例，统计归纳了三种形态的黏结漏钢时空特征，揭示了漏钢空间传播和动态演化规律，最后，采用决策树方法，对图像中的感兴趣区域特征进行快速提取，以感兴趣区域的热区域和冷区域特征为输入，建立支持向量机、随机森林智能化模型。研究结果表明：低合金钢比低碳钢更容易发生黏结漏钢，黏结漏钢实例中液位波动超过20mm的全部为低合金钢，结晶器内存在典型黏结漏钢、单侧传播黏结漏钢和大跨度黏结漏钢三种形态，其中，单侧传播黏结漏钢具有最大的升温速率和最低的降温速率，黏结漏钢具有典型的热区域和冷区域，黏结漏钢的热区域和冷区域面积、温度速率和位置特征的组合更适合作为判断真伪黏结漏钢的依据，以真伪黏结漏钢空间-时序特征为输入，建立了支持向量机和随机森林模型，漏钢预测准确率为93.8%。.截止到2020年12月31日，本项目研究已发表学术论文13篇，其中，SCI检索论文11篇，EI检索论文2篇，出版学术专著1部，授权发明专利1项，审查中发明专利1项，授权软件著作权1项。.项目研究对深刻认识结晶器内黏结漏钢形成和演化规律，提高连铸漏钢预报准确率，提升连铸结晶器在线监控的可视化、智能化水平具有重要的科学意义，在连铸工艺参数优化、漏钢预报等方面具有较好的应用前景。