中文摘要：

针对间歇过程特点和基于多向主元分析（Multiway Principal Component Analysis,MPCA）的间歇过程监控方法的缺陷,利用核映射在处理非线性过程和Fisher判别分析（Fisher Discriminant Analysis,FDA）在故障诊断能力上的优势,提出了基于递推多模型的核多向Fisher判别式分析（Recursive Multi-model Kernel Multi-way FDA,RMKMFDA）的间歇过程监测与故障诊断方法。该方法采用多模型核多向Fisher判别分析（Multi-model Kernel Multi-way FDA,MKMFDA）非线性结构代替MPCA单模型线性化结构,并提出确定时滞变量的算法;一旦通过MKMFDA监测出某一新批次过程正常,则模型参考数据库就随之更新;在线监控时通过比较核Fisher特征向量之间的欧氏距离来实现,而最优核Fisher判别向量用来鉴别故障类型。该方法在实时监控新的批过程时,只需利用已收集到的数据信息,且在线递推地更新模型参考数据库,提高了间歇过程监控的准确性,克服了MPCA不能处理非线性过程和实时性问题。通过采用RMKMFDA与移动窗多向主元分析（Moving Window MPCA,MWMPCA）方法对青霉素分批补料发酵过程的实时监控,结果表明RMKMFDA比MWMPCA能更及时地监测出过程异常情况,更准确地判断异常发生的原因。

英文摘要：

In view of the characteristics of batch process and the defect of batch process monitoring method based on multiway principal component analysis（MPCA）, using the advantage of kernel mapping in dealing with nonlinear process and the advantage of fishe r discriminant analysis（FDA） in the ability of fault diagnosis, a novel batch performance monitoring and fault diagnosis method based on recursive multi-model kernel multi-way FDA（RMKMFDA） was proposed. Multi-model kernel multi-way FDA（MKMFDA） instead of single model was used and how to calculate the time-lagged variable was proposed; Whenever a new batch detected by MKMFDA successfully remained within the bounds of normal operation, its batch data were added to the historical database of normal data and a new MKMFDA model was developed based on the revised database; The key to the proposed approach was to calculate the distance of block data which were projec ted to the optimal kernel Fisher discriminant vector between new batch and reference batch. Similar degree between the present disc riminant vector and the optimal discriminant vector of fault in historical data set was used to perform fault diagnosis. The proposed approach only uses the known data for on-line monitoring batch processes and can consecutively update model historical data set. The approach enhances the accuracy of batch process monitoring and overcomes nonlinear process and real-time problem which cannot be handled in MPCA. The proposed method was applied to monitoring fed-batch penicillin production, and the results clearly showed that, in comparison to the moving window MPCA method, the proposed method was more accurate and efficient to detect and diagnose the malfunctions.