Abstract: When the traditional multi-way partial least squares(MPLS)method was carried out for quality prediction, it performed the problems of low prediction accuracy and lack of local capacity. Considering to multiphase characteristics in industrial batch process, a mixture MPLS model method was proposed for the multiphases quality prediction. The Gauss mixture model(GMM)was employed to model the high dimensional spatial distribution of the measurement and the quality variables,which was used to identity sub-phase data blocks in each batch. Due to the problem of inequality length in identical sub-phase, the dynamic time warping(DTW)algorithm corresponding to the maximum path length of time was carried out to synchronize each data block in same sub-phase, whats more, the sub-phase MPLS model was set up according to the variable expansion method. The Fisher discriminate analysis(FDA)was introduced to minimize the relationship among sub-phase blocks, and then the kernel density method was used to monitor phases switch online by estimating the probability density distribution of less relationship sub-phase blocks. The multiphases quality prediction was established by mixing several sub-phase MPLS models according to Bays principle. Furthermore, the result of quality prediction for penicillin fermentation process showed the effectiveness of proposed method.

Key words: multi phase characteristics, bays principle, multi-phase model mixture, multi-way partial least squares, fisher discriminate analysis