GMM-DPLS间歇过程故障监测与质量预报

Fault monitoring and quality prediction for batch process using GMM-DPLS

工业间歇过程数据普遍具有多阶段、动态和非高斯特性,且轨迹不同步是其固有特征,针对上述问题,提出一种基于高斯混合模型-动态偏最小二乘(GMM-DPLS)的故障监测与质量预报新策略。采用GMM对过程数据进行聚类,客观反映不同阶段操作模态的数据分布特点,实现子阶段划分;针对子阶段不等长问题,采用动态时间规整(DTW)算法同步阶段轨迹,最后对同步后的子阶段分别建立DPLS模型。间歇发酵过程的应用实例表明该策略相比传统单一模型的DPLS方法,能显著提高故障监测效率和质量预报准确性。

In industrial manufacturing, most batch processes are inherently multiphase, dynamic, non-Gaussian behaviors and uneven-length batch processes in nature. To solve the aforementioned problem, a new strategy is proposed base on Gaussian mixture model（GMM）-dynamic partial least squares（DPLS） for batch process monitoring and quality prediction. Using GMM clustering arithmetic, batch process data was divided into several operation stages, since GMM is adopted to discriminate different operation modes. Then, run-to-ran variations among different instances of a phase are synchronized by using dynamic time warping （DTW）, and sub-phase DPLS models were developed for every phases. At last, the proposed method was applied to batch fermentation process. The results demonstrate that the more efficiency and accuracy of fault detection and quality prediction compared to the traditional DPLS.