Example 1 Detection of Abnormal Situations

This example illustrates a system for monitoring and detection in a continuous polymer process (Piovoso et al., 1992b). In the first stage of the process, several chemical reactions occur that produce a viscous polymer 

Such infrequent measurements make the control of the product quality difficult. At best, the operators have learned a set of heuristics that, if adhered to, usually produces a good product. However, unforeseen disturbances and undetected equipment degradations not accounted for by the heuristics still occur and affect the product. In addition, there are periods of operation when the final process step produces a degraded product in spite of near-perfect upstream operations. 

To address this situation, a data interpretation system was constructed to monitor and detect changes in the second stage that will significantly affect the product quality. It is here that critical properties are imparted to the process material. Intuitively, if the second stage can be monitored to anticipate shifts in normal process operation or to detect equipment failure, then corrective action can be taken to minimize these effects on the final product. One of the limitations of this approach is that disturbances that may affect the final product may not manifest themselves in the variables used to develop the reference model. The converse is also true—that disturbances in the monitored variables may not affect the final product. However, faced with few choices, the use of a reference model using the process data is a rational approach to monitor and to detect unusual process behavior, to improve process understanding, and to maintain continuous operation. 

We can expect throughput to have a major effect on all the measure- 

To construct the reference model, the interpretation system required routine process data collected over a period of several months. Cross-validation was applied to detect and remove outliers. Only data corresponding to normal process operations (that is, when top-grade product is made) were used in the model development. As stated earlier, the system ultimately involved two analysis approaches, both reduced-order models that capture dominant directions of variability in the data. A PLS analysis using two loadings explained about 60% of the variance in the measurements. A subsequent PCA analysis on the residuals showed that five principal components explain 90% of the residual variability. 

---