55 batches, each aligned to 100 equal time intervals (K = 100) for J=10 process variables measured throughout each batch. The batches are stacked vertically (rows 1-100 = batch 1, rows 101-200 = batch 2, etc). Values are scaled/normalized (dimensionless, roughly 0-1.8).
Variables 1, 2, 3: reactor temperatures
variables 6 and 7 are heating/cooling-medium temperatures
variables 4, 8, and 9 are pressures
variables 5 and 10 are flowrates of materials added to the reactor.
Known abnormal batches (from the paper)
Batches 50-55 stand out clearly in the score (t) plot.
Batch 49 stands out in the residual (Q) statistic; its quality was barely acceptable.
Batches 40, 41, 42, 50, 51, 53, 54, 55 had the final quality measurement well outside the acceptable limit;
Batches 38, 45, 46, 49, 52 were above or very close to the limit.
Batch 45 is independently known to have given bad product.
The paper builds its normal-operation reference model from 36 batches after excluding problematic ones.
Process
A two-stage batch polymerization.
Stage 1: ingredients are loaded; reactor-heating-medium flows are adjusted to control pressure and the rate of temperature change; the solvent used to convey ingredients is vaporized and removed. Vaporization is vigorous enough that the contents need no stirring.
Stage 2: the ingredients complete their reaction to yield the final polymer, again under controlled vessel pressure and temperature ramp. The batch finishes by pumping the polymer product out of the vessel.
Data source
The data were supplied by the DuPont Company from an industrial batch polymerization reactor and are the worked example in several publications. For example
Nomikos and MacGregor: "Multivariate SPC charts for monitoring batch processes", Technometrics, 37, 41-59, 1995. https://dx.doi.org/10.2307/1269152