Abstract:

Industrial process plants are often characterized with problems of variable drifts,nonlinearity and time-variant. The time difference (TD) model was proposed by researchers to handle the drifting problems. However, the global model used under TD model cannot describe the data characteristic like the time-variant and high nonlinearity well. Moreover, the prediction accuracy will greatly decrease when change of process state occurs. In this paper, the time difference model and locally weighted partial least squares (LWPLS) are synthesized to enhance the adaptability of soft sensor models. In the TD-LWPLS based soft sensor framework, TD is used to reduce the effect of process drifts. Moreover, as a just-in-time (JITL) method, LWPLS is utilized to tackle nonlinearity and change of process state. A numerical example and an industrial application example have been carried out to test the effectiveness and feasibility of the proposed method. The results demonstrate that the TD technique with the LWPLS model can achieve the best prediction accuracy in both cases compared to two other methods.

Key words: time difference model (TD), locally weighted partial least squares (LWPLS), just-in-time learning (JITL), soft sensor modeling, quality prediction

摘要：

工业过程软测量模型常常因为过程的变量漂移、非线性和时变等问题而使得预测性能下降。因此,时间差分已被应用于解决过程变量漂移问题。但是,时间差分框架下的全局模型往往不能很好地描述过程非线性和时变等特性。为此,提出了一种融合时间差分模型和局部加权偏最小二乘算法的自适应软测量建模方法。时间差分模型可以大大减少过程变量漂移的影响,而局部加权偏最小二乘算法作为一种即时学习方法,可以有效解决过程非线性和时变问题。该方法的有效性在数值例子和工业过程实例中得到了有效验证。

关键词: 时间差分模型, 局部加权偏最小二乘算法, 即时学习, 软测量建模, 质量预测