Batch processes have been widely employed in chemical and other manufacturing industries, due to their capabilities to meet the requirements of fast changing markets and to manufacture high-value-added products. In batch manufacturing, operation safety and product quality consistency are pivotal problems to guarantee high productivity, arousing attentions of many researchers. To deal with these two problems, many multivariate statistical analysis techniques such as principal component analysis (PCA) and partial least squares (PLS) have been adopted in the field of multivariate statistical process control (MSPC), and have been extended to different forms by taking the natures of batch processes into consideration. The related methods include multiway PCA/PLS (MPCA/MPLS), phase-ba...[ Read more ]

Batch processes have been widely employed in chemical and other manufacturing industries, due to their capabilities to meet the requirements of fast changing markets and to manufacture high-value-added products. In batch manufacturing, operation safety and product quality consistency are pivotal problems to guarantee high productivity, arousing attentions of many researchers. To deal with these two problems, many multivariate statistical analysis techniques such as principal component analysis (PCA) and partial least squares (PLS) have been adopted in the field of multivariate statistical process control (MSPC), and have been extended to different forms by taking the natures of batch processes into consideration. The related methods include multiway PCA/PLS (MPCA/MPLS), phase-based PCA/PLS, etc.

In the time direction within each batch, many existing methods assume that the similar characteristics within each phase can be captured by a single statistical model, which ignore the process variations between phases or within a phase. In the batch direction throughout the whole process, traditional methods handle process variations by adjusting the monitoring models consecutively in a direct way, which leads to the increase in probability of introducing disturbances and faults. In fact, process evolution, a typical kind of process characteristic variation caused by process dynamics or long term external factors, should be discriminated from the random variation caused by noises.

In this thesis, several novel strategies have been proposed to improve both batch process monitoring and quality prediction by tracing both inner-batch and inter-batch evolutions. Using the proposed methods, the specific issues contained in batch process data such as transitions, uneven durations, time-varying behaviors and multiple modes can be handled effectively.