Abstract: Accurate endpoint prediction of Basic Oxygen Furnace (BOF) carbon content and temperature is the key to realize endpoint control.To solve the problem of low quality of similar samples by traditional Just-In-Time Learning (JITL) measurement due to the high volatility and nonlinear characteristics of BOF process data,an adaptive JITL soft-sensing method based on Weighted Gaussian Kernel Self-Organization Map Dynamic Relevant Component Analysis (WGKSOM-DRCA) was proposed for endpoint prediction of BOF carbon content and temperature.The WGKSOM clustering algorithm was proposed by using the WGK metric criterion introducing label information to guide the clustering direction and improve the clustering quality of algorithm and reduce the influence of data volatility.The Gaussian posterior probability was used to calculate the membership degree of the test samples and the appropriate learning set was selected adaptively to predict the endpoint carbon content and temperature by introducing dynamic factors and using DRCA metric learning strategy.Results showed that the proposed algorithm performed better than other algorithms in predicting endpoint carbon content and temperature of BOF steelmaking.The prediction accuracy of carbon content was 92% within the error range of ±0.02%,and the prediction accuracy of temperature was 93.5% within the error range of ±10℃.

Key words: basic oxygen furnace, just-in-time-learning, soft sensor, self organization map, Gaussian kernel function, relevant component analysis

摘要： 转炉炼钢终点碳温的准确预测是实现转炉终点控制的关键。针对转炉生产过程数据存在波动性大和非线性特点引起传统即时学习度量的算法学习集质量低,进而削弱模型预测性能的问题,提出了一种基于加权高斯核自组织映射动态相关成分分析(WGKSOM-DRCA)自适应即时学习软测量建模方法用于转炉炼钢终点碳温预测。首先,采用引入标签信息的WGK度量准则构造WGKSOM聚类算法引导聚类方向,提高算法的聚类质量并降低类簇数据波动性对于建模的影响;其次,利用高斯后验概率计算待测样本的隶属度并通过引入动态因子构建DRCA度量策略,从而实现自适应的样本选择,进一步提升待测样本对应的局部算法学习集质量并用于局部模型训练,最终输出终点碳温的预测结果。实验表明,所提算法在转炉炼钢终点碳温预测上相对于其他算法有更好的表现,在±0.02%的预测误差范围,碳含量的预测精度为92%,在±10℃的误差范围,温度的预测精度为93.5%。

关键词: 转炉炼钢, 即时学习, 软测量, 自组织映射, 高斯核函数, 相关成分分析