shewhart

Statistical process control for Python.

Control charts with the standard run rules, process capability analysis, and measurement systems analysis. Results are computed from the published formulas and checked against reference values in the test suite.

Named after Walter A. Shewhart.

Motivation

R has had a maintained SPC package (qcc) since 2004. Python does not: the existing packages are unmaintained, cover only fragments of the toolkit, and none of them validate their output against reference data. This library is an attempt to fix that, with a few specific goals:

• correct constants and estimators, validated against published values
• a clean separation between estimating control limits (Phase I) and monitoring new data against frozen limits (Phase II)
• rule violations as structured data, usable in pipelines, not only in plots
• an API that works headless, so a weekly control chart review can run as a cron job

pip install shewhart

Documentation: https://bertanucar.github.io/shewhart/

Runnable examples: examples/

Status

Version 0.1.1 is on PyPI. Implemented and tested:

• review(): one call that selects the right chart, checks the assumptions, and returns a structured verdict (see below)
• control charts: I-MR, Xbar-R, Xbar-S, p, np, c, u (stair-step limits for varying subgroup sizes), Laney p'/u' for overdispersed data, EWMA (exact and asymptotic limits), tabular CUSUM, run chart with the four runs tests, Pareto analysis
• time-window subgrouping on DatetimeIndex data (subgroup="1H")
• process capability: Cp, Cpk, Pp, Ppk, Cpm with confidence intervals (chi-square for Cp/Pp, Bissell approximation for Cpk/Ppk), within vs overall sigma, expected and observed PPM, stability gate, normality check; non-normal data via fitted models (percentile method) or Box-Cox
• tolerance intervals: normal (Howe k2, anchored to the NIST handbook factor) and nonparametric (Wilks)
• named sigma estimators: average or median moving range, Sbar, pooled
• Nelson rules 1 to 8 and Western Electric rules 1 to 4, returned as structured signal events
• chart constants (d2, d3, c4, A2, A3, D3, D4, B3, B4), computed from their defining integrals rather than copied from tables
• baseline freezing and reuse (JSON), self-contained HTML reports
• a reference-case validation suite (tests/validation_cases.json), anchored externally: NIST StRD certified values (Michelso, NumAcc1) and the NIST/SEMATECH e-Handbook EWMA worked example are reproduced in CI

Usage

One call, if you just want a verdict:

import shewhart as sw

rv = sw.review(df, value="torque", lsl=9.95, usl=10.05)
rv.ok          # in control, capable, no failed checks
rv.headline    # "In control: no rule violations on the imr chart. Cpk 1.41 (capable)."
rv.to_dict()   # the full verdict as JSON-safe data

review() selects the chart from the data shape, checks the assumptions, and refuses to report capability on an unstable process. The individual analyses behind it:

r = sw.imr(df, value="torque", rules="nelson")
r.ok           # False if any rule fired
r.summary()    # plain text verdict
r.table        # per-point DataFrame with signal flags
r.plot()

Subgrouped data:

r = sw.xbar_r(df, value="torque", subgroup="batch")

Fit limits once, then monitor new data against them:

sw.imr(df_baseline, value="torque").baseline.save("line3_baseline.json")

# later, e.g. in a scheduled job:
r = sw.imr(df_new, value="torque", limits="line3_baseline.json")
sys.exit(0 if r.ok else 1)

Capability analysis, with the confidence intervals that the usual hand-rolled Cpk calculation cannot give you:

r = sw.capability(df, value="dia", lsl=9.95, usl=10.05)
r.stats["cpk"], r.stats["cpk_lci"], r.stats["cpk_uci"]

Several analyses in one self-contained HTML file, e.g. as a weekly job:

sw.report([
    sw.imr(df, value="torque", limits="line3_baseline.json"),
    sw.p_chart(df2, defectives="rejects", size="inspected"),
    sw.capability(df, value="torque", lsl=9.5, usl=11.0),
], "weekly_report.html", title="Line 3 weekly")

Every analysis returns the same Result object: named statistics, a tidy per-point table, a tuple of structured rule violations, and provenance metadata (library version, input hash, timestamp).

If you are wiring this into an AI agent, read Statistics is not a language task first.

Roadmap

Version	Scope
0.2	measurement systems analysis: ANOVA gauge R&R (crossed and nested), Type 1 studies, attribute agreement
0.3	process screening across many characteristics, drift monitoring with control chart semantics

Out of scope: DOE (see pyDOE3), reliability engineering (see reliability), general statistics (see statsmodels), GUIs.

License

MIT. Written and maintained by Bertan Ucar, PhD researcher at Tsinghua University.