| Summary: | 碩士 === 中原大學 === 化學工程研究所 === 90 ===
Batch production processes play an important role in chemical industries. Pharmaceuticals, biochemicals, semiconductors and polymers, for example, often utilize batch production. Batch processes are characterized by prescribed processing of raw materials into products within finite duration. The profiles as the fingerprint of the batch operations provide vital information characteristics of the operation of the batch. They can be used to identify if the current operating condition is successful. In this research, an online batch process monitoring based on the three-way data analysis is developed because the data in the batch operation are usually arranged in a three-way matrix with batch, measurements and their time profiles. Based on parallel factor analysis (PARAFAC), the developed technique extracts the state of the system via applications of mathematical and statistical methods from the big volume of the past historical database.
Before the on-line PARAFAC model is developed, a systematic derivation explains why PARAFAC is chosen over MPCA. The variance of the abnormal data based on the estimated model is derived. It is not necessary to carry out the derivation for the abnormal profile estimates. The variance of the abnormal data can be expressed into two Jacobians terms to capture the amount of the abnormal variations. The issues of the detection of the abnormal variance suffered from the number of components are presented. From this analysis, it can clearly indicate PARAFAC is more robust than MPCA. Then on-line batch monitoring methods, referred to as DPARFAC and Tri-DPLS, are developed. They integrate the time-lagged windows of the process dynamic behavior with PARAFAC and tri-PLS respectively. They deal with dynamic relationships; that is, the measured variables at one time instant have the serial correlation within variable series at the past time instances. Like our previously developed BDPCA and BDPLS, DPARAFAC and Tri-DPLS models only collect the previous data during the batch run without expensive computations to anticipate the future measurements. This leads to simple monitoring charts, easy tracking of the progress in each batch run and monitoring the occurrence of the observable upsets. Several examples are used to investigate the potential applications of the proposed methods and make a comparison with the previous on-line methods.
|
|---|
