贝叶斯模型比较的多模型组合框架在软测量建模中的应用

针对基于单一模型建立的软测量模型存在着预测精度需要进一步提高的问题,在分析目前常用的2种多模型组合框架的基础上,提出了一种基于贝叶斯模型比较的多模型组合框架。该框架以通过模糊c-均值聚类分析获得的生产过程状态变化知识为基础,对每种状态下各子模型的预测性能采用贝叶斯模型比较方法进行比较,并以此为基础在不同状态下采用了不同的子模型加权策略。在进行模型比较时,基于交叉检验分布,使用子模型训练所得采样序列,有效地减少了计算量。将该框架用于工程应用,取得了较好效果。

A multi model composition framework based on bayesian model comparision and its application in soft sensor modeling

In order to improve the prediction performance of single model based soft sensor, the features of the current model combination frameworksby analynizing, a new multi model combination framework based on the bayesian model comparison is proposed. In this framework, fuzzy c means clustering to the historial data is used to analyze the production states, then the prediction performance of sub models at different states are compared based on bayesian model comparison. The comparing results are the basis of the model combination stratery at different states. With adapting cross validation predictive distribution, the samples got from the trained models are used to successfully reduce computation load of model comparion.The framework has obtained good results in the practical application.