基于SOM-PCA-RVM 的过程建模及其应用

针对呈现高度非线性,自变量之间存在交互作用,且采集的过程数据具有一定类别特征的复杂系统,提出基于自组织映射神经网络-主元分析-关联向量机相结合的建模方法。首先,通过SOM,将样本分割成模式特性相近的若干子类,实现样本模式空间的分割。然后,基于每一子空间的建模样本,提取主元,并以预测性能为指标确定最佳主元个数,消除冗余信息干扰。最后,将各子空间的主元分别作为RVM模型的输入,建立各自的模型,实现基于样本模式空间分割的分类建模。仿真试验和在精对苯二甲酸生产过程对羧基苯甲醛含量软测量中的实际应用表明,SOM-PCA-RVM模型的拟合精度和预测精度不仅优于RVM模型,也优于PCA-RVM模型。

Process Modeling Based on SOM-PCA-RVM and Its Application

A hybrid modeling method for the complex process was proposed. The process, which was highly nonlinearity, had interacting independent variables, and had process data belonged to different categories. This method combined self- organizing map network （SOM network）, principal component analysis （PCA）, and relevance vector machine （RVM）. First, divided the sample datum into several subspaces which had the same patterns with the utilizing of SOM, the pattern space separation was realized. Then derived the principal components based on the modeling samples of each subspace, defined the optimal principal component numbers based on prediction ability, and filtered the redundancy. At last, the principal component of each subspace was applied as the input of RVM, building the models individually, and the classifying modeling was realized based on the separated of sample pattern spaces. The simulation results and the application in 4--earboxybenzaldehyde content soft sensor of pureed terephthalic acid production show that the accuracies of regressing and predicting of SOM--PCA--RVM model is better than both RVM model and PCA--RVM model.