基于数据驱动多输出ARMAX建模的高炉十字测温中心温度在线估计

高炉（Blast furnace，BF）炼铁中，十字测温作为炉顶温度和煤气流分布监测的最主要手段，对高炉的安全、稳定和高效运行起着重要作用.然而，由于高炉炉顶中心部位温度较高，造成十字测温装置中心位置传感器极易损坏，并且更换周期长，因而无法及时判断炉顶煤气流分布.针对这一实际工程问题，本文基于时间序列建模思想，集成采用多输出自回归移动平均（Multi-output autoregressive moving average，M-ARMAX）建模、因子分析、Pearson相关分析、基于赤池信息准则（Akaike information criterion，AIC）与模型拟合优度联合定阶等混合技术，提出一种模型结构简单、精度较高且易于工程实现的十字测温中心温度在线估计方法.首先，提出利用因子分析与Pearson相关分析相结合的稳健特征选择方法选取多输出建模输入变量.然后，采用样本均值消去法预处理采集的高炉样本数据，使其成为离散随机数.基于离散随机数，建立算法简单、易于工程实现的M-ARMAX温度模型:为了克服传统基于AIC阶数确定造成模型阶次高、结构复杂的问题，提出在AIC准则基础上进一步引入模型拟合优度来选取模型最小阶，可保证模型估计精度的同时降低模型阶次；同时，采用可快速收敛的递推最小二乘算法辨识M-ARMAX模型参数，并用残差分析方法检验模型.工业试验和比较分析表明:建立的M-ARMAX模型能够根据实时数据同时对十字测温装置多个中心温度点进行准确和稳定估计，且模型估计误差符合高斯白噪声特性.

Data-driven Multi-output ARMAX Modeling for Online Estimation of Central Temperatures for Cross Temperature Measuring in Blast Furnace Ironmaking

In a blast furnace (BF) ironmaking process, cross temperature measuring is the most important means for monitoring the temperature and gas flow distribution of furnace top. It plays an important role in safe, stable and efficient operation of the whole BF. However, due to the high temperature in the middle of furnace top, the central position sensors of the cross temperature measurement are easily damaged, and the replacement period always takes a long time, thus the gas flow distribution cannot be monitored in time. To solve such a practical engineering problem, a data-driven model with simple structure and high estimation precision is proposed for online estimation of central temperatures for cross temperature measuring. This model integrates hybrid modeling technologies, such as multi-output autoregressive moving average (M-ARMAX) modeling, factor analysis, Pearson correlation analysis, co-determination of model order by Akaike information criterion (AIC) and goodness-of-fit evaluation, etc. The input variables are determined by using factor analysis combined with Pearson correlation analysis with robustness, and the sample mean elimination method is used to preprocess the BF data to make them become discrete time random data. Discrete random data based M-ARMAX modeling includes determination of model order and identification model parameters. The conventional AIC based model order determination leads to too high order of model structure. Thus, the model goodness-of-fit is introduced in the AIC to select the minimum model order, which can not only guarantee the model accuracy but also reduce the model order. Finally, the parameters of M-ARMAX model are identified by recursive least squares algorithm and the obtained models are verified using residual analysis method. Industrial tests and comparative analysis show that the established M-ARMAX model can accurately estimate the center temperatures of cross temperature measuring devices, whose estimated error conforms with Gauss white noise characteristics.