A CV inspection model ships to the line, returns a clean pass/fail verdict on every unit, and the team treats that verdict as the whole quality story. It is not. The verdict stream is raw data, and raw data on a production line belongs inside statistical process control before anyone acts on it. Read a single reject and you react to one unit. Plot the defect rate against control limits and you can tell whether the process moved, whether the model moved, or whether nothing moved at all and you are chasing noise. That distinction is the entire point of pairing statistical process control tools with a computer-vision defect-detection station. SPC is the operations-side instrument that converts a noisy per-unit verdict into an actionable signal: defect rate against control limits, false-reject rate, and time-to-detect a process shift. It does not replace model monitoring — it makes drift legible to the people who already own the line, in a language they already speak. How Do Statistical Process Control Tools Work When the Signal Comes from a CV Model? The classic SPC toolkit — control charts, run charts, Pareto analysis, subgroup sampling, capability indices — was built for measured gauge values: a bore diameter, a fill weight, a torque reading. Each measurement carries inherent variation, and SPC exists to answer one question repeatedly: is the variation I am seeing the normal common-cause noise of a stable process, or is it a special-cause signal that something changed? Nothing in that question requires the signal to come from a caliper. A CV inspection station produces a stream of verdicts, and from that stream you derive the same statistics quality engineers have charted for decades. The natural primitive for attribute inspection — pass or fail per unit — is the p-chart (proportion defective) or the np-chart (count defective in a fixed subgroup). If the model emits a continuous score rather than a hard verdict — a defect-area estimate, a confidence margin, a measured gap width — you can run a variables chart (X-bar and R, or an individuals-moving-range chart) on that continuous output instead, which is statistically richer and detects drift earlier. The reframe that matters: the CV model is not the quality system. It is a measurement instrument feeding a quality system. Treat it the way you would treat any new gauge on the line — characterise it, baseline it, then chart its output. The same logic carries through our broader account of how CV defect-detection models survive the move from pilot to production; SPC is the layer that catches what model-side monitoring alone is slow to see. Which SPC Tools and Charts Apply to CV Defect-Detection Output? Not every tool in the classic “7 QC tools” set earns a place line-side, and the ones that do play distinct roles. The table below maps the common SPC instruments to what they actually buy you when the input is CV verdicts rather than gauge readings. SPC tool What it charts from CV output What it catches Evidence class p-chart / np-chart Defect proportion per subgroup of units Process shift in true defect rate; out-of-control runs observed-pattern Individuals-MR chart A per-unit continuous score (defect area, confidence) Early drift before binary verdicts flip; single-unit excursions observed-pattern False-reject sub-chart Reject rate confirmed good by audit re-check Model over-rejecting — the silent scrap driver observed-pattern Pareto chart Defect-class frequency (model’s class labels) Which defect type dominates; where to focus root cause observed-pattern Cp / Cpk thinking Spread of continuous output vs spec limits Whether the station has the resolution to discriminate at all observed-pattern A pairing worth singling out: run the defect-rate p-chart and a false-reject sub-chart side by side. A genuine process shift pushes the true defect rate out of its control limits; a model drift or a lighting change often shows up first as a rising false-reject rate while the true defect rate sits flat. Watching only the headline reject count conflates the two, and the two demand opposite responses — one is a process problem for operations, the other is a model problem for the CV team. For the chart mechanics and signal rules in depth, our companion piece on SPC control charts for CV defect-detection goes deeper than this overview does. How Do You Set Control Limits When the Verdict Comes from a Model? Control limits are not specification limits — that confusion causes more bad SPC than any other single mistake. Specification limits come from the customer or the design; control limits come from the process’s own demonstrated behaviour. You compute them from a stable baseline window, conventionally at plus or minus three sigma around the centre line. When the measurement instrument is a CV model, the baseline window has to be earned, not assumed. Before you trust the station’s output inside a control chart, you need two things characterised: First, measurement reliability — the CV equivalent of a Gauge R&R study. Feed the station a set of known-good and known-defect units, repeatedly, under the lighting and fixturing the line actually runs. If the model returns inconsistent verdicts on the same physical unit, its measurement variation is eating into your process variation and your control limits will be meaningless. In configurations we have characterised, an inspection station that cannot repeat its own verdict on a re-presented unit is not ready to anchor a control chart, regardless of its headline accuracy on a clean test set. Second, a baseline period that reflects normal operation — typically several shifts spanning the real run of part variants, operators, and ambient conditions, not a single calibrated demonstration hour. Compute the centre line and limits from that window, freeze them, and treat the pilot baseline as the reference the line will later be compared against. When the live chart diverges from that frozen baseline, that divergence is itself a documented trigger — which is where the rollback discussion later in this article begins. The same Cp/Cpk thinking from gauge work applies: if the spread of the model’s continuous output cannot cleanly separate good from defective relative to the spec, no amount of charting fixes it. That is a capability problem in the inspection station, and it belongs to the feasibility conversation, not the SPC conversation — see when industrial CV inspection actually works for where that line gets drawn. How Can SPC Charts Act as an Early Warning That the Model Is Drifting? This is the move that earns SPC its place alongside model monitoring rather than as a substitute for it. The control chart you built to watch the process doubles as a sensor on the instrument watching the process. The mechanism is straightforward once you see it. A CV model’s accuracy rarely collapses overnight; it erodes — a packaging supplier changes a label gloss, a fixture wears, the morning sun hits a window it never used to. Model-side monitoring that watches prediction confidence can be slow to register this because the model stays confident while being confidently wrong. The control chart, by contrast, registers the consequence directly: the false-reject rate creeps toward its upper control limit, or the defect-rate chart shows a run of points on one side of the centre line. Western Electric and Nelson rules — runs of consecutive points above the mean, trends, points beyond two sigma — were designed exactly to catch a slow drift before any single point breaches the three-sigma limit. So the same Nelson-rule run that, on a gauge chart, says “the process is drifting” says, on a CV-fed chart, “either the process drifted or the model did — go look.” The chart cannot tell you which. What it gives you is a timestamped, operator-visible early signal that something changed, hours or shifts before quietly regressing accuracy would have shown up as escaped defects or accumulated scrap. A line that watches raw verdicts alone reacts late, after the cost has already landed. That early-warning property is precisely the instrumentation our production-ai-monitoring-harness work specifies for industrial CV drift detection. When Should an Out-of-Control Signal Trigger Re-Calibration or Rollback? An out-of-control signal is a question, not a verdict. The discipline is in the response protocol, decided before the signal ever fires. A useful rubric: Confirm the signal is real. A single point beyond three sigma can be a measurement artefact. A Nelson-rule run — say, eight consecutive points on one side of the centre line, or a sustained upward trend — is far harder to dismiss. Do not act on one excursion; act on a pattern. Separate the two suspects. Pull a sample of the units the chart is reacting to and re-inspect them by the audit method (manual, or a trusted secondary station). If the units are genuinely defective, the process shifted — that is an operations problem. If the units are good and the model is rejecting them, the model drifted — that is a CV problem. Map the response to the suspect. A confirmed process shift triggers the line’s normal corrective action. A confirmed model drift triggers re-calibration or, if the divergence from the frozen pilot baseline exceeds the agreed threshold, rollback to the last validated model version. Document the trigger. The control-chart divergence, the audit re-check result, and the decision become the record. That documented trigger is what makes “the model is drifting” an auditable event rather than an argument. The threshold for rollback should be set against the pilot baseline, not invented in the moment. When control charts diverge from that baseline by the agreed margin, the rollback path is already defined and the line does not have to negotiate it under pressure. Who Owns the Charts Versus the Model, and How Do They Hand Off? The honest answer is that two teams share one instrument, and the hand-off is where most of these deployments fray. The quality team owns the SPC charts — they read them every shift, they call the out-of-control signals, they speak the language fluently. The CV team owns the model — its training data, its versioning, its re-calibration. The chart is the shared boundary object. When the quality team’s chart flags a run, the audit re-check in step 2 above is the formal hand-off: it routes the signal to operations if the units are bad and to the CV team if the model is wrong. Get that re-check protocol agreed in writing before go-live and the two teams cooperate; leave it implicit and every out-of-control signal becomes a turf argument about whose problem it is. This is the same reliability-artefact discipline a hardened deployment produces and signs against, covered in our account of the artefacts that keep a line-side CV model running, and it sits inside the broader computer vision practice we bring to industrial inspection. When the question becomes how to scope that hardening as an engagement, that is where how we work with manufacturing teams starts. FAQ How does statistical process control tools work, and what does it mean in practice? SPC tools — control charts, run charts, Pareto analysis, subgroup sampling, capability indices — answer one repeated question: is the variation I see normal common-cause noise from a stable process, or a special-cause signal that something changed. In practice on a CV-inspected line, the model’s verdict stream becomes the input, and the chart tells the operator whether to react or to leave a stable process alone. What SPC tools and charts apply to CV defect-detection output on a production line? For binary pass/fail verdicts, the p-chart and np-chart track defect proportion or count per subgroup. If the model emits a continuous score, an individuals-moving-range chart on that score detects drift earlier. A false-reject sub-chart, a Pareto of defect classes, and Cp/Cpk thinking on the continuous output round out the set, each catching a distinct failure. How do you set control limits when the inspection verdict comes from a CV model rather than a manual gauge? Control limits come from the process’s own demonstrated behaviour over a stable baseline window — conventionally three sigma around the centre line — not from specification limits. First characterise the station’s measurement reliability with a Gauge R&R-style repeatability check, then compute limits from a baseline spanning real part variants, shifts, and ambient conditions. Freeze that baseline as the reference the live chart is later compared against. How can SPC control charts act as an early warning that the CV model is drifting? A model’s accuracy erodes slowly, and confidence-based monitoring can stay confident while being wrong. The control chart registers the consequence directly: a rising false-reject rate or a Nelson-rule run of points on one side of the centre line. That timestamped, operator-visible signal often fires hours or shifts before quietly regressing accuracy shows up as escaped defects or scrap. How do you separate normal process variation from a genuine process shift in defect-rate data? Control limits define the band of normal common-cause variation; points inside the limits with no run pattern are noise you should not react to. A genuine shift shows as points beyond three sigma or as Western Electric / Nelson rule patterns — runs, trends, points beyond two sigma. Acting only on patterns, not single excursions, is what stops the line from chasing noise. When should an out-of-control SPC signal trigger model re-calibration or rollback? Confirm the signal is a pattern rather than a single artefact, then re-inspect the flagged units by an audit method. If the units are genuinely defective the process shifted; if they are good and the model rejected them the model drifted. A confirmed drift triggers re-calibration, or rollback to the last validated model version when divergence from the frozen pilot baseline exceeds the agreed threshold. Who owns the SPC charts versus the CV model in production, and how do they hand off? The quality team owns and reads the SPC charts; the CV team owns the model, its versioning, and re-calibration. The chart is the shared boundary object, and the audit re-check is the formal hand-off — it routes a flagged signal to operations if the units are bad and to the CV team if the model is wrong. Agreeing that re-check protocol in writing before go-live prevents every out-of-control signal from becoming a turf argument.
