Does on-line SPC require real-time automation?

No. Periodic manual review can be on-line when it supports ongoing process monitoring and decisions.

Does a signal prove a special cause exists?

No. It is evidence warranting investigation and can occur by chance under the monitoring rule.

Is a process within control limits necessarily good enough?

No. It may be stable while failing practical or specification requirements.

AI Glossary / Statistics and probability / Intermediate

On-line SPC

Definition

On-line SPC means statistical process control used to monitor an operating process and identify when its behavior warrants investigation or corrective action relative to an established pattern of variation. Control charts are a common example. Some approaches also support gradual process optimization during operation. Here on-line describes the connection between statistical monitoring and process decisions; it does not require an internet connection, instantaneous computation, or automatic intervention after every unusual observation.

What it means

Every process varies. On-line SPC aims to distinguish ordinary variation compatible with the established process from evidence that something may have changed. A signal prompts investigation; it does not by itself identify the cause or prove that output is unacceptable.

Monitoring preserves the order of observations, because the timing and persistence of changes matter. A process can produce the same overall histogram while showing very different time patterns. Effective monitoring therefore connects the statistical display with knowledge of the process and an agreed response procedure.

How it works

Define the output measure, sampling plan, and operational decisions. Establish a credible baseline using a sufficiently understood process and appropriate measurement system. Select a chart suited to the data, such as subgroup means, individual measurements, proportions, or counts.

Calculate limits and signal rules using assumptions appropriate to that chart. As new observations arrive, compare the chosen statistic with the established limits and inspect relevant patterns. Investigate signals using process records, equipment checks, and other evidence before deciding on corrective action.

Document findings and avoid routine adjustment for ordinary noise, which can increase variation. Reassess the baseline after a justified, sustained process change rather than continually recalculating limits to absorb deterioration. Autocorrelation, changing sample sizes, seasonality, and evolving case mix may require a model or chart that accounts for those features instead of applying a simple independent-observation rule blindly.

Three practical examples

A production team charts subgroup mean dimensions and investigates signals associated with possible equipment changes.
A service operation monitors a defined error proportion over successive periods, accounting for the number of opportunities in each period.
An AI-supported workflow tracks a stable quality measure and routes unusual patterns to an owner for investigation rather than silently modifying the system.

Think of it this way

Think of monitoring a vehicle's instrument panel against its normal operating behavior. A warning tells the operator to investigate, but the cause could be a sensor problem, a temporary condition, or a mechanical fault. A useful monitoring system needs both a meaningful signal and a sensible response.

Worked example

Suppose a fictional stable process has known mean 100 units and known individual standard deviation 2 units. Every review uses the mean of four independent, approximately normal measurements. The subgroup-mean standard error is 2/sqrt(4) = 1 unit.

Illustrative three-standard-error limits are therefore 100 − 3 = 97 and 100 + 3 = 103. A new subgroup mean of 104 lies four standard errors above the baseline mean and exceeds the upper limit, producing a signal under this rule.

These are control limits for subgroup means, not product specification limits or limits for individual measurements. The example assumes known baseline quantities and independent observations. In practice, estimating the baseline and monitoring many times affect operating characteristics, and a signal still requires investigation before attributing a cause or changing the process.

Illustrative case study

Fictional case study: Cedar Service monitors the accuracy of a repetitive document-processing workflow. A new chart shows a sustained change after an input-format update. The operations team checks whether the change reflects actual errors, a revised scoring rule, or a different mix of documents.

After identifying and addressing the relevant issue, it documents the intervention and gathers evidence before establishing a revised baseline. The dashboard records both the statistical signal and the investigation outcome. This fictional case illustrates monitoring discipline; it does not claim that a chart alone diagnoses problems or guarantees reliable AI performance.

Common misconceptions and mistakes

A common mistake is equating control limits with customer requirements. A stable process can consistently fail specifications, while an unstable process may temporarily remain within them. Stability and acceptability are separate questions.

Another mistake is adjusting the process after every small fluctuation. This can respond to noise rather than a meaningful change. Adding many signal rules increases opportunities for false alarms, and repeated observations are not automatically independent. A chart also cannot establish that a model is fair, safe, or useful on every relevant dimension merely because one monitored statistic appears stable.

Business use

On-line SPC helps organizations detect changes in operational quality, prioritize investigations, and assess whether improvements persist. It is useful when the process and measurement definitions are sufficiently consistent for a baseline comparison to make sense.

AI can assist with data collection, annotation, or investigation summaries, but monitoring needs accountable ownership and documented response rules. Track changes in data definitions and workload composition. Where the process is highly adaptive, assess whether the monitoring assumptions remain valid rather than interpreting every change as either failure or improvement.

Check your understanding

A subgroup-mean chart has limits of 97 and 103. Should those limits automatically be applied to each individual observation?

Answer: No. The limits describe the variability of subgroup means under the chart's assumptions. Individual observations have a different variability scale and require an appropriate monitoring method.

Related terms

Statistical process control
Control chart
Off-line SPC
Control limits
Rational subgroups
common cause variation
special cause variation
process stability
rational subgroup
false alarm

Learn next

Learn how common and special causes are distinguished and how subgrouping affects chart behavior. Then compare chart types and study false-alarm rates, detection speed, and response procedures. Practice evaluating stability separately from capability and business acceptability.

Sources

The NIST Engineering Statistics Handbook glossary defines on-line SPC: https://www.itl.nist.gov/div898/handbook/glossary.htm

Its control-chart introduction explains monitoring concepts and limits: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc31.htm

The original numerical example uses known baseline parameters for clarity. It is not a prescription for setting limits from a small operational sample.

Previous: On-Demand Features Next: On-policy