面向转炉炼钢终点控制的火焰吹炼信息特征提取与熔池碳温连续实时预报模型研究

The accurate endpoint control is an important prerequisite of improving converter steel quality and reducing steelmaking energy consumption. The key and bottleneck of the endpoint control is to achieve a non-contact, continuous and real-time carbon temperature prediction at present. The flame blowing information features can be used to realize the continuous detection and fuse with additional data to predict output in real time. The purpose of this study is to propose a more accurate method of extracting flame blowing information features and molten pool carbon and temperature detection model. Firstly, a quaternion multi-scale convolution neural network model is proposed. Multi-scale four-element convolution and weighted fusion are implemented in one single channel and in a whole way, which helps to extract flame non-structural depth features and avoid information loss. Secondly, a multi-scale three-dimensional convolutional neural network model is proposed to extract sequence dynamic deformation features. Flame deformation sequence are operated multi-scale three-dimensional convolution in space-time range, which used to extract dynamic characteristics of the flame deformation and overcome the instability combustion disturbance deformation. Finally, a molten pool carbon temperature detection model is proposed based on feature selection and membership weighted cost function, and fused of steelmaking process data and flame-blowing information key features. The model has the anti-jamming performance in the process of automatic training. The project research methods and conclusions have great significance to the converter endpoint automatic control, improvement of steelmaking quality and efficiency, and energy conservation.

吹炼终点的准确控制是提高转炉炼钢质量和降低冶炼能耗的重要前提，目前终点控制的关键和瓶颈在于碳含量和温度的非接触连续实时预报。火焰吹炼信息特征可实现连续检测并融合加料数据实时调整预报输出，本研究旨在构建更符合火焰特点的吹炼信息特征提取方法和准确的熔池碳温预报模型。首先，提出四元数多尺度卷积神经网络模型，在单通道内以整体方式实现多尺度四元卷积和加权融合处理，提取火焰非结构性深度特征并避免信息丢失；其次，提出一种多尺度立体卷积神经网络序列动态形变特征提取模型，以火焰形变序列为输入在时空范围进行多尺度立体卷积，提取火焰动态形变语义特征并克服不稳定燃烧等干扰形变；最后，提出一种特征选择和隶属度加权代价函数的熔池碳温预报模型，融合了炼钢生产过程数据和火焰吹炼信息关键特征，并使模型在自动训练过程中具有抗干扰性能。项目研究所用方法和结论对转炉终点自动控制、提高炼钢质量和效率、节能减排等方面具有重要的意义。

转炉炼钢终点准确控制是提高钢铁质量和实现节能减排的关键，而其难点在于实现对吹炼过程中熔池内碳含量和温度的准确连续实时预报。本研究从提取转炉炼钢过程火焰图像吹炼信息特征的准确表述以及构建基于火焰图像识别的终点碳含量预报模型、以及采用转炉炼钢过程数据处理与分析，构建过程数据驱动的转炉终点碳温连续预报模型方法。主要研究内容、结果及科学意义如下：.针对转炉火焰图像存在的多尺度、随机性和多方向性等问题，传统特征提取方法难以获得准确的吹炼信息特征表述，研究火焰特征本身所具有的特点，设计并提出了多种非刚体类对象的图像特征表征算法并应用于转炉炼钢终点碳含量连续实时预报。通过提出四元数复杂网络、导数非线性映射方向加权多层复杂网络等方法适应于火焰图像不规则、多方向和多尺度等特点，能够准确表述火焰图像随着吹炼过程而呈现的特征变化规律。在此基础上，构建了火焰图像特征到吹炼终点碳含量之间的连续预报模型，测试实验结果表明，所提方法在准确度和实时性等方面能够较好满足工业实际需求。此外，为了更精确实现转炉终点碳含量和温度的准确测量，同时对火焰图像信息特征提取的终点预报进行互补融合，进一步提升终点碳温预报的鲁棒性和准确性，研究转炉炼钢过程数据的降维和特征选择理论，设计并提出了多种基于深度学习、精确相似性度量的即时学习以及集成学习的转炉终点碳温预报模型，并获得了满意的测试结果，方法的研究丰富和扩展了在高维度、非线性和不稳定等过程数据背景下的软测量建模理论，并为转炉炼钢终点碳温准确实时预报提供了重要研究基础和结论。.通过本项目的研究，为转炉炼钢吹炼终点碳温连续实时预报提供了重要研究方案和实现思路，扩展和丰富了图像特征表述以及软测量等领域的研究方法，为进一步实现转炉炼钢自动化生产和终点精确控制提供了基础和依据。

内容获取失败，请点击重试