基于CJS-SLLE降维与即时学习的转炉炼钢终点碳温软测量方法

Soft sensor method of endpoint carbon content and temperature of BOF steelmaking based on CJS-SLLE and just-in-time learning

中文关键词  转炉炼钢  即时学习  相似性度量  预测分析  降维

英文关键词  BOF steelmaking  JITL  similarity measure  predictive analytics  dimensionality reduction

中文摘要

      转炉炼钢中碳温的准确检测是终点判断的关键, 基于数据驱动的终点碳温软测量方法是一种有效途径, 但转炉炼钢生产过程数据存在高维度、非线性和数据波动大的问题. 针对这一问题, 本文提出一种降维与即时学习的终点碳温软测量(CJS-SLLE)算法用于过程数据的监督降维. 通过在距离度量中引入量化后的碳温标签信息, 从而构造了一种带有监督信息的度量方式实现类内类间方差的调整, 然后在带标签信息的基础上引入数据间方向信息, 从而实现了样本标签、方向和距离三者信息融合的一种新型(CJS)相似性度量策略, 应用到局部线性嵌入中获得高维训练样本低维坐标; 其次, 提出一种自适应局部线性投影策略用于无标签待测样本, 实现其低维坐标中同样包含标签信息; 最后, 根据即时学习算法选取样本子集建立偏最小二乘局部回归模型对终点碳温预测. 在实际转炉炼钢生产过程数据仿真下, 碳含量在±0.02%误差范围内预测精度达到90%, 温度在±10℃误差范围内预测精度达到87%

英文摘要

      Accurate detection of carbon temperature in basic oxygen furnace (BOF) steelmaking is the key to end point judgment, and the data-driven end point carbon temperature soft sensor method is an effective way, but the data in BOF steelmaking production process has problems of high dimension, nonlinearity and large data fluctuation. In response to this problem, a Cosine similarity and Jensen Shannon divergence supervised local linear embedding (CJS-SLLE) dimensionality reduction algorithm is proposed for supervised dimensionality reduction of process data. By introducing the quantified carbon temperature label information into the distance measurement, a measurement method with supervision information is constructed to adjust the variance between classes within a class. Then the direction information between data is introduced on the basis of the label information, so as to achieve A novel CJS similarity measurement strategy based on the fusion of sample label, direction and distance information is combined with local linear embedding to obtain low-dimensional coordinates of training samples; Secondly, a new method is proposed. The adaptive local linear projection strategy is used for unlabeled samples to be tested. Their low-dimensional coordinates also contain label information; Finally, a partial least squares local regression model is established to predict the endpoint carbon temperature by selecting a subset of samples according to the just-in-time learning algorithm. Under the data simulation of the actual BOF steelmaking production process, the carbon content has an accuracy of 90% within the error range of ±0.02%, and the temperature has an accuracy of 87% within the error range of ±10℃.