00241 tbbf0afd5fc79c2ad0c3ecbce2b5261

Statistical Process Monitoring

243

0

20

40

60

Batch Number

0

100

Figurę 2. Plot of viscosity (corrected for changes in catalyst) from 109 batches of polymer resin. The downward shift around batch 89 might have been preventable and it should have been detected by a monitoring scheme. A standard control chart with widened control limits, however, would not have detected the shift.

shift occurs around batch 89. If this shift stands out statistically from the typical wandering naturę of the process, it might point to a cause of variation that could be removed. For example, a shift could result from changing the feedstock from one siło to another or from a sudden drop in the ambient temperaturę at the plant. If the cause of a shift is found it may be possible to remove it or at least compensate for it as in the case of weather related causes.

In this example, it is important to detect a sudden shift in yiscosity of the same order of magnitude as the typical swings of the process level. Simply widening the usual limits on a standard control chart would help remove unwanted signals but that would also severely limit the chart’s ability to detect unexpected abrupt changes in the process such as the one that may have occurred at batch 89. A different solution is needed here and this paper provides some guidance.

Monitoring forecast errors One approach to detecting a sudden shift in a wandering process uses forecast errors from a time series model. If