局部自适应加权LSSVM在线建模方法及其在间歇过程中的应用

On-line modeling based on local adaptive weighted LSSVM and its application in the batch process

目前的局部建模方法在构建样本间相似度的时候仅考虑了输入信息而忽略了输出信息的作用,并且没有考虑样本的权重问题。针对上述问题,提出了局部自适应加权最小二乘支持向量机(Local Adaptive Weight LSSVM,LAW-LSSVM)回归算法。该算法采用同时考虑输入输出信息的相似性判据则来构建更加合理的相似样本集,利用有监督的局部保持映射(Supervised LocalityPreserving Projection,SLPP)算法对样本空间进行有效的降维和搜索最优的相似样本方向,实现了样本权重的在线调整。利用LAW-LSSVM对青霉素发酵过程中的产物浓度进行在线预测,仿真结果表明,包含了输入输出信息的相似度评价准则能够更准确的选择相似样本,较离线LSSVM以及局部LSSVM(LLSSVM)有着更高的预测精度、更好的泛化能力。

Compared with global learning approaches, local learning has a better accuracy and generalization ability.However, local learning methods always only utilize input information to select relevant instances, which may lead to a waste of output information and inaccurate sample selection. What is more, it ignores that different sample has different weight, which can affect the accuracy of the modeling. To overcome these disadvantages, a new local modeling algorithm, local adaptive weight LSSVM （LAW-LSSVM） is proposed, in which both input and output information are used in a new similarity measurement, and a supervised locality preserving projection technique is utilized to select relevant samples. In LAW-LSSVM, instead of using traditional cross-validation methods, the trade-off parameters are adjusted iteratively and the local model is updated recursively, which reduces the computational complexity a lot. The proposed LAW-LSSVM is applied to the online prediction of biomass concentration in the penicillin fed-batch process. The simulations showed that LAW-LSSVM could predict the biomass concentration online accurately, and the information contained in the output information is conducive to choose the fight similar samples.Compared with LSSVM offline and normal local LSSVM, the proposed LAW-LSSVM algorithm has better generalization ability.