基于进化多目标优化的选择性集成学习软测量建模

Selective ensemble learning based on evolutionary multi-objective optimization for soft sensor development

常规集成学习软测量方法忽略了输入变量选择的多样性,而且没有对基模型进行修剪,从而造成集成模型复杂度高、预测性能受限。为此,提出一种基于进化多目标优化(EMO)的选择性集成学习(SE)高斯过程回归(GPR)软测量建模方法,称为 EMO-SEGPR。该方法融合输入特征扰动,通过结合 bootstrapping随机重采样和偏互信息相关分析(PMI)构建多样性输入变量子集,并据此建立多样性 GPR 基模型。然后,基于 EMO 算法对 GPR 基模型进行集成修剪,从而获得一组集成规模较小、多样性和准确性较高的基模型。最后,引入集成学习策略实现 GPR 基模型的融合。将EMO-SEGPR 方法应用于青霉素发酵过程和 Tennessee Eastman 化工过程,实验结果表明了该方法的有效性和优越性。

Traditional ensemble soft sensors usually ignore the diversity of input variable selection and simply combine all base models without pruning, which may result in high model complexity and poor prediction performance. An evolutionary multi-objective optimization (EMO) based selective ensemble Gaussian process regression (GPR) model, referred to as EMO-SEGPR, was proposed for soft sensor development. The method employed the input feature perturbation to build diverse input variable sets by combining bootstrapping resampling and partial mutual information (PMI) based relevance analysis, and then diverse base GPR models were constructed. Furthermore, these GPR models were pruned by an EMO based ensemble pruning approach to generate a set of base models with small ensemble size, high accuracy and high diversity. Finally, the selected base GPR models were integrated using ensemble methods. The effectiveness and superiority of EMO-SEGPR were verified through penicillin fermentation process and Tennessee Eastman chemical process.