基于实时学习的高斯过程回归多模型融合建模

Multi-model Combination Modeling Based on Just-in-time Learning Using Gaussian Process Regression

为提高软测量模型的预测性能,降低化工过程中的非线性、多阶段性和不同的局部动态特性对产品质量的控制产生的影响,提出一种在线不断更新的多模型策略.该方法用高斯混合模型(GMM)对过程的不同阶段进行辨识,并采用一种自适应实时学习(JITL)方法,不断更新所建立的高斯过程回归(GPR)模型.当新的数据到来时,在每个不同的阶段,基于欧氏距离和角度原则选择部分相似的数据,用于建立局部的高斯过程回归模型.最终根据计算得到的新的数据隶属于每个不同阶段的后验概率,对局部模型进行融合输出.与传统的单个模型相比,这种实时学习软测量模型的结构更加灵活,而且能更好地跟踪过程的动态.基于常用的TE(Tennessee Eastman)化工过程,利用本方法对产品的质量进行预测,仿真结果表明了所提方法具有更高的预测精度和更好的泛化性能.

To improve the prediction performance of soft sensor models and to reduce the influence on product quality caused by various challenges including process nonlinearity, multiple operating phases, and different local dy- namics, we propose a multi-model soft sensor method based on the just-in-time learning （JITL） method. The proposed method uses the Gaussian mixture model （GMM） to distinguish the data from different operating pha- ses and applies an adaptive JITL strategy to update the built models. The relation between input and output data is modeled using the Gaussian process regression （GPR） model. Whenever a new sample is available, local GPR models are constructed using a portion of the most relevant samples selected by the Euclidean distance and angle method in the different operating phases. Then, according to the posterior probabilities of the sample be- longing to the different operating phases, the predictions of the local GPR models are combined to obtain the de- sired global output. Compared with traditional soft sensors based on a single model, the JITL-based approach exhibits a more flexible structure and the process dynamics can be captured better. A Tennessee Eastman （TE） chemical process is used to demonstrate the feasibility and effectiveness of the proposed approach. The results show that the proposed approach provides higher predictive accuracy and better generation ability.