Copilot Plotting Kit — Plate Gallery

IMatplotlib conventions

Templates and annotation helpers every static figure is built from. Object-oriented API, preset sizes, direct labeling, semantic reference bands.

Plate C-01 fig_templates.py::stacked

Aligned subplot stack. Three quantities, one shared time axis, one physical unit per axes. Warn and fail bands shaded with semantic tokens; the key moment carries a callout rather than a legend entry.

Plate C-02 fig_templates.py::comparison_bars

Comparison bars. Highlight-one, grey-the-rest: the proposed detector in house blue, alternatives in neutral. Values labeled at the bar tip, zero baseline, no vertical gridlines.

IIDiagnostics plots

Fault detection and isolation visuals: spectra, time-frequency maps, residuals against thresholds, and classifier evaluation under class imbalance.

Plate D-01 spectral.py::plot_spectrum

Welch spectrum with fault markers. dB magnitude with the expected 6th-harmonic location marked as a labeled stem — hypotheses on the plot, not in the legend.

Plate D-02 spectral.py::plot_spectrogram

Spectrogram. Sequential cream-to-blue colormap on the light canvas; the event of interest framed by an ink box. Mesh rasterized inside the SVG to keep files small.

Plate D-03 spectral.py::plot_residual_threshold

Residual and threshold. The three mandatory layers: residual, dashed threshold, alarm flags — plus the fault-injection window shaded and first-alarm time called out. Detection delay is THE number.

Plate D-04 classifier_eval.py::plot_confusion

Confusion matrix. Raw count and row-normalized recall in every cell; diagonal readable at a glance, text color flips for legibility on dark cells.

Plate D-05 classifier_eval.py::plot_roc_pr

ROC and precision-recall pair. Always shown as a pair — at 6% fault prevalence an excellent ROC can hide useless precision. The deployed operating point is marked on both.

Plate D-06 fault_heatmap.py::plot_signature_matrix

FDI signature matrix. Ternary fault-symptom structure: up/down/none glyphs, the isolability of each fault mode visible by row pattern.

Plate D-07 fault_heatmap.py::plot_dtc_occurrence

DTC occurrence heatmap. Fleet DTC counts over weeks; the peak cell auto-framed and annotated — the figure points at its own anomaly.

IIIPrognostics plots

Remaining useful life, degradation, and reliability. Uncertainty is the product: no bare point estimates, ever.

Plate P-01 rul_fan.py::plot_rul_fan

RUL fan. Median trajectory with nested 50% and 90% bands against ground truth; end-of-life is the dashed crimson floor.

Plate P-02 rul_fan.py::plot_alpha_lambda

Alpha-lambda accuracy. The standard PHM accuracy view: predictions inside the ±20% cone in blue, violations in crimson, in-cone fraction graded in the title.

Plate P-03 degradation.py::plot_health_index

Health index with P-F interval. Green-amber-red zones as quiet background shading; the P-F window — the maintenance opportunity, and the business case — measured and annotated.

Plate P-04 degradation.py::plot_fleet_spaghetti

Fleet spaghetti. Twenty-five units in neutral grey with a percentile band; one unit of interest in blue. Narrative view of “this vehicle versus the fleet.”

Plate P-05 survival.py::plot_weibull_survival

Weibull survival. Empirical steps with the fitted curve; B10 and B50 lives dropped to the axes with shape and scale reported in the legend.

Plate P-06 survival.py::plot_hazard

Hazard rate. The bathtub, with infant, useful-life, and wear-out phases labeled in place.

IVTime-series telemetry

Vehicle logs at scale. This 400,000-sample log is min-max decimated to ~4,000 points per signal — note the single-sample 140 A spike at t ≈ 90 s that naive striding would have erased.

Plate T-01 multisignal.py::plot_signal_stack

Decimated signal stack. Aligned stack with a mode lane (enum states as a step plot, not numeric lines) and a DTC event marker spanning all axes. The preserved spike is the proof of the decimator.

VDistributions & SPC

Population comparisons and drift monitoring: the shift-readable ECDF, sorted violins, ridgeline drift, Q-Q honesty checks, and EWMA/CUSUM control charts for fleet features.

Plate X-01 dist_plots.py::plot_ecdf

ECDF comparison. The whole-population answer to “did the fix work?” — every percentile of the after-fix curve sits left of before; P90 marked for the spec conversation.

Plate X-02 dist_plots.py::plot_violin_panel

Violin panel. Groups sorted by median with n under each violin and the overall median as a dashed reference — plant C is genuinely tighter, plant D genuinely worse.

Plate X-03 dist_plots.py::plot_ridgeline

Ridgeline drift. Distribution evolution across five software releases, oldest at the bottom; the tail above 4 A visibly disappears.

Plate X-04 dist_plots.py::plot_qq

Q-Q normality check. Heavy tails depart from the fitted line at both ends — a Gaussian threshold on these residuals would over-alarm. The honesty companion to any threshold figure.

Plate X-05 control_charts.py::plot_ewma

EWMA control chart. A 0.9σ sustained shift, invisible in the raw samples, signaled 13 points after onset; first violation called out, rule parameters in the title.

Plate X-06 control_charts.py::plot_cusum

CUSUM chart. Both one-sided sums against the decision interval h — the upward sum climbs steadily after the shift while the downward stays flat.

VIMaps & 2D fields

The native visuals of machine calibration and operating analysis — and the home of the kit’s 3D policy: surfaces only where the surface IS the object, always with a contour twin.

Plate M-01 efficiency_map.py::plot_efficiency_map

Efficiency map. Iso-efficiency lines labeled at the business-relevant levels, torque envelope in ink, MTPA→MTPV trajectory direct-labeled, drive-cycle points sized by dwell.

Plate M-02 flux_maps.py::plot_flux_surface

Flux map surface. Policy-compliant 3D: ψd over the current plane with saturation droop toward the viewer, contour projection on the floor, measurement grid as light edge lines.

Plate M-03 flux_maps.py::plot_flux_contour

Contour twin. Same data, measurable form — reviewers read values here. The twin is mandatory for any 3D figure that ships.

Plate M-04 dwell_density.py::plot_dwell_hexbin

Dwell density. 80,000 fleet operating points as a log-scaled hexbin against the capability envelope; the dashed crimson box flags a region the fleet uses but testing doesn’t cover.

VIIML insights

Model understanding for fault classifiers and fleet segmentation — embeddings that state their method, clustering claims with their silhouette honesty check, and calibration for probabilities that feed maintenance decisions.

Plate L-01 embeddings.py::plot_embedding_clusters

Embedding clusters. PCA scatter with convex hulls and per-cluster n; axes labeled with explained variance so the projection is falsifiable.

Plate L-02 embeddings.py::plot_silhouette

Silhouette plot. Per-cluster coefficient profiles against the mean — the mandatory companion to any clustering claim; thin or negative clusters are visible immediately.

Plate L-03 correlation.py::plot_corr_clustered

Clustered correlation. Features reordered by hierarchical clustering so correlated blocks emerge — the electrical-signature block and the thermal block separate cleanly; weak cells washed out.

Plate L-04 importance.py::plot_feature_importance

Feature importance. Sorted bars with cross-fold spread, top-5 in blue, the rest neutral; the method is named because gain, permutation, and SHAP rank differently.

Plate L-05 importance.py::plot_calibration

Calibration diagram. Reliability curve with ECE self-graded in the title and the prediction histogram as a quiet strip — for probabilities driving maintenance, calibration beats AUC.

VIIIFlows & relations

How totals decompose, volumes flow, and faults propagate. The waterfall is the canonical executive ROI exhibit; the DAG renderer refuses cycles by design.

Plate F-01 waterfall.py::plot_waterfall

ROI waterfall. Gains in green largest-first, costs in crimson, net anchored in slate — the one-figure version of a program pitch. Exec-polish checklist applies by default.

Plate F-02 dag_graph.py::plot_layered_dag

Layered causal DAG. Root causes → symptoms → effects, layered by causal depth; edge width encodes occurrence. Accent semantics match the rest of the kit.

Plate F-03 parallel_coords.py::plot_parallel

Parallel coordinates. Failed units in blue against a subsampled grey fleet across five normalized axes with real-unit endpoints — the failure population runs hotter, harder, and charges faster.

IXSVG diagrams

Conceptual figures — built from grid-snapped primitives, text stays selectable text. Accent semantics match the charts: blue plant, teal estimation, orange evaluation, crimson decision.

Plate S-01 svg_primitives.py::Diagram

Model-based FDI loop. Blocks, labeled arrows, and consistent color meaning. Data charts get matplotlib; structure gets this.

XTables and KPI cards

Live HTML, not screenshots — the same tokens as the charts. Status cells use tinted fills with a colored left border; one heat-shaded column ranks the table.

Plate TB-01 styled_tables.py::kpi_cards · df_to_html_table

Fleet coverage

94%

+6 pts vs Q1

False alarms

2.1/Mh

−38%

Avoided downtime

$1.4M

+$0.3M

Table 1 — Detector performance by fault mode (validation fleet).
Fault mode	Detection rate	FAR [1/Mh]	Median horizon [h]	Status
Bearing wear	0.96	1.20	310	Pass
Stator ITSC	0.91	3.40	120	Pass
Resolver offset	0.99	0.40	540	Pass
DC-link cap	0.83	6.80	45	Review

Executive KPI strip and detector table. Seven-row ceiling, right-aligned tabular numerals, em-dash for missing values, caption under the table.

XIInteractive (plotly)

These plates degrade gracefully: a static SVG twin renders everywhere (including offline file previews); when the page is online and plotly.js loads, each upgrades in place to the fully interactive chart — hover, zoom, unified tooltips.

Plate I-01 plotly_theme.py::timeseries · add_threshold

Interactive telemetry. Unified hover with units, overspeed limit as a labeled semantic line. Shown as its static twin until plotly.js loads, then upgraded in place.

Plate I-02 sankey_flow.py::build_sankey

DTC resolution Sankey. Stage-colored nodes, links at 30% alpha of their source, volumes in the hover. Flows under 2% of total are grouped before plotting.

Plate I-03 fleet_geo.py::plot_state_choropleth

Fleet DTC rate choropleth. Rates normalized by vehicles in operation — never raw counts; states under the minimum-n threshold greyed out. The cold-climate pattern is the hypothesis on display.