改进的FCM-LSSVM青霉素发酵过程预测建模

Improved FCM-LSSVM Prediction Model for Penicillin Fed-batch Fermentation

针对青霉素发酵过程周期长,每个阶段表现出不同的特性,最小二乘支持向量机(least squares support vector machines,LSSVM)的全局模型预测精度难以保证的问题,提出了改进的基于模糊C均值聚类(fuzzy C-means clustering algorithm,FCM)和LSSVM的青霉素发酵过程分段建模方法。首先,在分析影响青霉素产物浓度相关因素的基础上选取输入变量,对样本数据采用FCM算法聚类,按照最大隶属度将样本归类为稳定过程或过渡过程;然后,分别为稳定过程的4个阶段和过渡过程的3个阶段分别建立LSSVM子模型,最后通过子模型切换策略得到系统输出。利用Pensim仿真平台数据,将提出的方法与FCM-LSSVM和LSSVM方法进行比较,平均绝对误差分别为0.013 2、0.014 3、0.014 9,均方根误差分别为0.017 8、0.019 2、0.021 6,实验结果表明,所提出的方法具有良好的精度和泛化能力。

Due to the long cycle of penicillin fed-batch fermentation process which has different characters in each stage, it is difficult to guarantee the prediction accuracy of global modeling by using single least squares support vector machines （LSSVM）. To solve this problem, an improved staged modeling method based on fuzzy C-means clustering algorithm （FCM） and multi-LSSVM is proposed. Firstly, the factors which affect the penicillin concentration most are analyzed and selected as the inputs, then, samples are classified to stabilization or transition process by using FCM algorithm; Secondly, sub-models of LSSVM are built for disparate stages of stabilization or transition process. Finally, the system output is obtained through the sub-model switch strategy. Using the simulation data from the Pensim simulation platform and utilizing mean absolute error and root mean squares error indexes to analyze the performance of the proposed method, FCM-LSSVM and LSSVM, and mean absolute errors are 0.013 2, 0.014 3, 0.014 9 respectively, root mean squares errors are 0.017 8, 0.019 2, 0.021 6 respectively, simulation results illustrate the effectiveness of the proposed method.