摘要
酒精发酵的pH值具有非线性、时变性和动态性。利用常规辩识方法对pH值进行辩识,一方面,无法准确描述其动态特性;另一方面,由于常规神经网络的权值学习是梯度下降法,在训练过程易陷入局部极小,并且训练速度慢。针对这些问题,将改进的动态递归神经网络应用于pH值的辩识研究。通过实验验证了该算法不但能体现出发酵过程的动态特性,而且通过在动态递归神经网络的权值学习中引入滤波项,能有效地克服常规网络在权值学习过程中的问题。表明该算法对pH值辩识的有效性。
Alcohol fermentation process is a non-linear and time-varying dynamic process.On the one hand,modeling of pH with conventional identification methods is impossible to accurately describe the dynamic characteristics.On the other hand,weight learning method of conventional neural network,the gradient descent method,is easy to fall into local minima,and is slow in the training process.To solve these problems,pH modeling fermentation process based on improved dynamic recurrent neural networks is studied.Experiments show the validity of the algorithm in pH modeling.Filtering is introduced to the dynamic recurrent neural networks to learn the weight of the recurrent neural networks.It can solve the problems in the weight learning of conventional network effcetively.Experiments show the validity of the algorithm in pH modeling.
出处
《控制工程》
CSCD
北大核心
2009年第S2期83-86,共4页
Control Engineering of China
基金
上海师范大学重点学科基金资助项目(DZL811)
上海教委科研创新重点基金资助项目(09ZZ141)
上海师范大学前瞻性基金资助项目(DYL200809)
国家自然科学基金资助项目(60572055)
关键词
动态递归神经网络
酒精发酵
PH
滤波
dynamic recurrent neural networks
alcoholic fermentation
pH
filtering
