丁二酸发酵过程的软测量建模研究

Soft sensor modeling study of succinic acid fermentation processes

生物转化法生产丁二酸的间歇厌氧发酵过程存在明显的不确定性和高度非线性,其中某些参数(如丁二酸浓度)难以在线检测,给过程优化策略的有效实施带来了障碍。最小二乘支持向量机(LS-SVM)是标准支持向量机(SVM)的一种扩展,遵循结构风险最小化原则。将该算法用于丁二酸发酵过程建模,用具有RBF核函数的LS-SVM建立丁二酸浓度的模型,并通过MATLAB 7.0开发工具和径向基(RBF)人工神经网络的建模方法进行比较。结果表明LS-SVM方法比基于RBF神经网络的软测量建模方法降低了83.7%的外推误差,具有更好的泛化能力,使针对丁二酸浓度的在线预估与优化控制成为可能。

Many critical variable, are difficult to be obtained in succinic acid fermentation process, which holds back the effective application of optimal procedure strategy. Least square support vector machine （LS-SVM） is a kind of extension of support vector machine （SVM） which obeys structural risk minimization （SRM） during training. It has been widely used in biochemical process modeling recently. LS-SVM was introduced to model succinic acid fermentation process. A model was built by LS-SVM with RBF kernel for succinic acid concentration. Finally, the modeling effect was analyzed through toolbox in MATLAB 7.0 and compared with the model which was built by radial basis function （RBF） artificial neural network. The result shows that the algorithm of LS-SVM has high precision and better generalization ability than RBF neural network. The LS-SVM model greatly reduces the extrapolation distance error, it indicates that LS-SVM is an effective method for succinic acid fermentation process modeling, predicting and controlled real-time and on-line.