摘要
为实现发酵过程重要变量的预测,提出基于批次加权正则极限学习机的软测量模型。结合发酵过程中各批次变量变化轨迹与发酵初始条件密切相关的特点,采用欧式距离描述各训练批次初始条件与预测对象初始条件之间的相似度,设计了一种新的相似度量化函数求解各训练批次的惩罚权值,实现了批次加权正则极限学习机建模;另外,针对正则极限学习机中的超参数估计问题,采用贝叶斯方法对超参数进行估计,降低了计算代价且实现了参数自适应估计。将其应用于青霉素发酵过程产物质量浓度的软测量中,仿真结果表明该方法预测精度高,效果好。
The soft senor of fermentation processes based on the BWRELM is proposed to measure the significant variables.Considering the actual feature of the change track of the variables and the initial conditions of the fermentation process,this paper takes the Euclidean distance as the similarity between the initial conditions of each batch of the training samples and the forecast object,models the weighted RELM based similarity quantitative function to gain the penalty weights for each batch of training samples.According to the parameter estimation in RELM,this paper estimates the hyper-parameters by using the Bayesian method which could reduce the computational cost and obtain the adaptive parameter estimation.Applying the proposed method to model the penicillin fermentation process,the simulation results are good.
出处
《江南大学学报（自然科学版）》
CAS
2013年第5期515-521,共7页
Joural of Jiangnan University (Natural Science Edition)
基金
国家自然科学基金项目(61134007)
江苏高等学校优秀科技创新团队项目
江苏高校优势学科建设工程项目
关键词
发酵过程
软测量
正则极限学习机
批次加权
贝叶斯参数估计
fermentation process
soft sensor
regularized extreme learning machine
batch weighting
Bayesian parameter estimation
