传统的偏最小二乘方法(partial least squares,PLS)因未对建模数据求取均值轨迹,以及没有考虑多单元生产对浸出率的综合作用,导致无法准确建立过程信息与质量变量之间的回归关系。根据高铜矿氰化浸出过程的多单元和耗时长的特点,提出一种针对连续过程的基于多单元均值轨迹的浸出率预测方法。获取建模数据的均值轨迹矩阵,在此基础上分别建立每个单元与实测浸出率的回归模型。定义输入向量与每个单元建模数据的相似度以及预测模型的权重,将各单元预测结果加权综合作为最终预测值。将该方法应用于氰化浸出过程浸出率预测,仿真结果表明,该方法体现了生产过程实际物理特性,提高了模型的解释能力,增强了预测模型的泛化性能。
针对诺西肽发酵过程中菌体质量浓度的估计问题,提出了一种基于RBF神经网络的软测量建模方法.在诺西肽发酵过程非结构模型的基础上,根据隐函数存在定理确定出辅助变量,从而使其选择有严格的理论依据;根据每批样本数据对被预测对象的预估能力,自适应地为各个批次的训练样本分配权值,并进而实施加权RBF神经网络建模.实际应用表明,所提出的软测量建模方法是有效的.
Abstract：
A RBF neural network based soft sensor method is presented for the estimation of biomass in Nnsiheptide fermentation process. Based on the unstructured model of Nosiheptide fermentation process, the secondary variables are selected according to the implicit function existence theorem, which makes the selection be strict in theory. Each batch training samples are self-adaptively weighted according to their different predicting ability to the predicted object, and then weighted RBF neural network (WRBFNN) is applied to develop the biomass soft sensor modeL The testing result shows that the presented method is effective.
