Figures for the DuPont, SBR and FMC batch case-study pages

[kgdunn]

Summary

• Adds the figures for three new batch case-study pages in the book's Applications chapter (Three batch case studies for the Applications chapter (DuPont batch PCA, SBR batch PLS, FMC multiblock batch PLS) pid-book#273): the DuPont batch PCA outlier hunt, the SBR batch PLS fault diagnosis, and the FMC multiblock batch PLS ladder.
• Three matplotlib scripts in batch/ (batch-case-dupont-figures.py, batch-case-sbr-figures.py, batch-case-fmc-figures.py) with a shared style module batch/batch_case_common.py, and the 34 PNGs they write (batch/batch-case-*.png). The scripts run the same analysis as the pages' Plotly code on process-improve 1.81 (Batch case studies: DuPont batch PCA, SBR batch PLS, FMC multiblock batch PLS (#154, #155, #156) process-improve#537: BatchPCA, BatchPLS, BatchMonitor, load_dupont, load_fmc, load_sbr, unfolded_contribution_plot, contributions for rows with missing cells, the r2_per_variable_ fix and the SCP score-estimator fix).
• Every score plot carries the percent of variance each component explains on its axes (explained_per_component in the shared module: R2 of X per component, from r2_per_variable_ for PLS-type models; R2Y on the multiblock super-score panels). The axes of every PLS-type score plot, block-score panels included, are labelled R2X, so that they are not read as the R2Y of the super scores.
• The DuPont model B score plot draws a thick arrow, labelled "contribution direction", from a square at the second group's average point to the model centre: the displacement the group contribution figure decomposes. That figure shows the per-tag bars with the eight members as white dots with a dark edge on the left, on alternately shaded positions, and on the right a 2x2 block of raw trajectories over the first 30 samples for the three tags with the largest contributions and for Flow-2. The model C figure projects the 15 batches left out of that model onto it and draws them on the T2-against-SPE panel, on logarithmic axes, beside the 40 training batches.
• The SBR score plot marks batch 4, the batch nearest the average quality, in purple, because the page's mid-batch prediction figure is drawn for it. The SBR script also writes the four figures of the page's two on-line sections: the leave-one-batch-out prediction error (RMSEP) per attribute against samples observed, batch 4's evolving prediction of particle size and composition with a band of one RMSEP at each sample against its final prediction and measured value (integer ticks on the particle-size axis), the on-line monitoring of batch 37 (Hotelling's T2) and batch 34 (SPE of the newest sample) with the residual shares at the alarm sample, and the trajectory forecast of the remainder of both faulty batches. The leave-one-out sweep (53 refits, traced at every sample) adds about a minute to the script.
• The SBR departure figure draws two signed distances of each faulty batch from the other batches per tag: a robust one (median and 1.4826 MAD of the others, EWMA-smoothed with lambda 0.3) as a solid line and the mean-and-standard-deviation one dashed.
• The FMC script carries ClockTime as the eleventh tag of the unfolded trajectory block, as the original study did, and the batch-13 overlay shows the clock time in place of the agitator power. Every FMC score plot is coded by the plant's disposition, a colour and a marker shape per class (blue circles good, purple triangles abnormal, gold squares high solvent, the squares drawn a little smaller so that the three markers read as the same size), at twice the side of a plain dot so that a triangle and a square can be told apart, with the highlighted batches larger still in their class's marker; legends with four entries sit in a 2 x 2 grid, and the block-scores figure has one legend row under its three panels. The bar panels shade alternate positions, and grids are drawn behind the bars and markers (axes.axisbelow). The multiblock PLS on the two initial-condition blocks is a 2 x 3 figure: the super scores in the first column, and for each block its scores above its weights (a weight_plot helper in the script), so that batch 20, inside the cloud in the chemistry block and far outside it in the operating-condition block, can be read against the weights that place it there. The loading panels of the batch PCA carry the cumulative R2 of every cell after two components on a second axis in faint orange. The two phase ends (samples 175 and 249) are marked as vertical lines on the raw overlays, the loading panels, the per-sample panel of batch 20's contribution figure and the raw panels of the four-batch figure. The SPE contribution figure is drawn for batch 20, the batch above both limits, now that process-improve computes contributions for a batch with missing cells: its missing samples 34 to 109 are blank positions and the dryer pressure carries half of the residual (it replaces the batch-41 figure). Two figures are new: the block scores of the batch multiblock PLS, one panel per block, with batches 2, 3, 6 and 7 labelled (classed good, placed with the abnormal batches by the trajectory block and with the good ones by both initial-condition blocks); and the operating-condition contribution from those batches' nearest abnormal neighbours to them, beside the raw collector level, clock time, dryer temperature and its set point for both groups.
• batch_case_common.py gains online_chart (one batch's per-sample statistic against the reference-batch mean and the dashed per-sample limit), group_scatter (per-class colour and marker, used by score_plot and influence_plot through their groups and group_styles options, with the marker areas MARKER, MARKER_CODED, HIGHLIGHT, HIGHLIGHT_CODED and a per-marker area scale), compact_legend (two columns from four entries), label_bars, shade_alternate_tags, secondary and vlines options on tag_panels, overlay_panels and contribution_triptych (a second right-hand axis and vertical lines at the phase ends), a label_left option on score_plot and parity_plot and a legend_loc option on score_plot and influence_plot (for the labels and legends that would otherwise sit on a neighbour), and three more categorical colours (purple, gold for the FMC high-solvent class, magenta for the SBR page).
• All three data sets are live on openmv.net when the scripts run (polymerization.csv, batch-dryer.xlsx, sbr-batch-reactor.xlsx); the SBR script keeps a --data-url option for a local copy when working offline.

Colours are the book's house colours (dark blue, orange, with aqua as the second highlight), checked as a categorical set with a colour-vision-deficiency validator; raw overlays and per-tag panels use small multiples rather than many-hued lines.

Library dependency of the FMC loading panels. The orange R2 curves read PCA.r2_per_variable_, which on the library's NIPALS path (taken because the trajectories have missing cells) used to hold the share of the remaining residual that each component explained rather than the cumulative R2; the committed PNG was regenerated against the fix in kgdunn/process-improve#537 (the collector-tank cells read 0.07 before and 0.78 after). Regenerating it against an older library reproduces the wrong curve.

Script checks. The "Against process-improve main" and "Against the PyPI release" jobs fail on the three batch scripts because they import load_dupont, load_fmc, load_sbr and BatchMonitor, which exist only on the kgdunn/process-improve#537 branch (the PyPI release does not even carry BatchPCA). They pass once #537 merges (main job) and a release carries it (PyPI job); nothing in this repository fixes them.

Merge order: this PR first, then the book PR kgdunn/pid-book#273, so the book's PDF build finds the images.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE

---

Generated by Claude Code

[kgdunn]

The two red script checks ("Against the PyPI release" and "Against process-improve main (advisory)") fail on the three batch case-study scripts only, at their imports: load_dupont, load_fmc, load_sbr and BatchMonitor exist only on the kgdunn/process-improve#537 branch, and the PyPI release does not carry BatchPCA at all. The other 23 scripts pass and 16 are skipped for the optional pyoptex dependency, as before.

Nothing in this repository can make those two jobs green: the main job passes once #537 merges, and the PyPI job once a release carries it. I am not changing anything here for them.

---

Generated by Claude Code