基于双层机器学习的动态精馏过程故障检测与分离  

Fault Detection and Isolation of Dynamic Distillation Process Using Two-tier Machine Learning

作　　者：毛海涛[1] 田文德[1] 梁慧婷[1] 

机构地区：[1]青岛科技大学化工学院,山东青岛266042

出　　处：《过程工程学报》2017年第2期351-356,共6页The Chinese Journal of Process Engineering

基　　金：国家自然科学基金资助项目(编号:21576143);山东省自然科学基金资助项目(编号:ZR2013BL008);国家级大学生创新创业训练计划资助项目(编号:201510426003)

摘　　要：提出了基于双层机器学习的动态精馏过程故障检测和分离的方法,检测的阈值为正常工况训练的网络输出值与样本的残差.通过对比网络预测值和实测值的偏差检测故障,检测到故障时,启动另一网络对动态过程自适应拟合异常工况数据.网络的预测值与实测值的偏差小于阈值时,拟合成功.通过对两个网络进行结构解析找到造成输出变量异常波动的输入变量.将该方法运用到脱丙烷精馏塔中,检测出过程中的故障,并分离出与故障源相关的变量,表明该方法准确、有效.A new method using two-tier machine learning is proposed to detect and isolate fault in dynamic distillation process. The residuals between output of network trained by normal condition data and samples are recognized as the threshold for detection. Fault detection is carried out by comparing the deviation between the prediction of one network and the measured value. When the trouble is detected, another network is activated to fit the dynamic distillation process adaptively. When the deviation between simulation output and the measured output of distillation column is less than the threshold, the fitting is considered satisfying. Then the input variables causing output variables＇ abnormal fluctuation are found via the analysis of structure parameters of two networks. This method is applied to detect process＇s fault and isolate variables relating with fault in the distillation tower simulation, and proved to be effective and veracious.

关 键 词：机器学习 动态精馏过程 故障检测与分离 网络结构解析 

分 类 号：TQ018[化学工程]