论文部分内容阅读
针对传统的多向偏最小二乘方法(multi-way partial least squares,MPLS)在质量预报中存在着模型预测精度低、局部预报能力不足等问题,提出一种多MPLS模型融合方法来提高预报表现。利用高斯混合模型(Gauss mixture model,GM M)对每批次过程和质量数据组成的高维空间进行阶段识别。针对多批次同一子阶段长度不等问题,采用动态时间规整(dynamic time warping,DTW)算法依据最长持续时间同步为等长轨迹,并在子阶段中按变量展开方式建立MPLS模型。根据Fisher判据分析(Fisher discriminate analysis,FDA)最小化子阶段数据集间相关性,利用核密度方法估计子阶段数据集去相关后的概率密度分布来在线监测阶段切换。利用贝叶斯原则融合各子阶段MPLS模型进行质量预报。将该方法应用到工业青霉素发酵过程中,表明了所提方法具有更好的监控性能和预报能力。
Traditional multi-way partial least squares (multi-way partial least squares) methods have some problems in the quality prediction, such as low prediction accuracy and insufficient local prediction ability. A multi-way partial least squares method is proposed to improve the prediction performance . The high-dimensional space composed of the process and quality data of each batch was identified by Gaussian mixture model (GMM). Aiming at the problem that the length of the same sub-stage is different from one batch to another, a dynamic time warping (DTW) algorithm is used to synchronize the same-length trajectory according to the longest duration and build the MPLS model by variable expansion in the sub-stage. According to the Fisher discriminate analysis (FDA), the correlation between the data sets of the sub-stage minimization was determined and the phase density was estimated by using the kernel density method to estimate the sub-stage data set after the correlation. The Bayesian principle is used to predict the quality of each sub-phase MPLS model. The method was applied to industrial penicillin fermentation process, indicating that the proposed method has better monitoring performance and forecasting ability.