Statistical Process Control and Improvement Sustainment

Introduction and Context

Lean manufacturing removes activities that do not create customer value, while Six Sigma reduces process variation and defects. Statistical Process Control (SPC) connects these aims by providing a disciplined method for distinguishing normal process behaviour from signals requiring investigation. SPC is not simply the production of charts. It is a management system for collecting data, interpreting process evidence, responding consistently, and learning from variation.

A process may be stable but incapable of meeting specifications, or apparently capable but unstable. These conditions require different actions. Control charts address stability; capability analysis addresses performance relative to customer or engineering requirements. Improvement sustainment then ensures that gains survive personnel changes, demand fluctuations, equipment deterioration, and competing operational priorities.

This chapter develops the selection and interpretation of variable and attribute control charts, control plans, reaction plans, standard work, dashboards, audits, and layered process reviews. It also examines ownership, training, governance, and change management as essential conditions for institutionalising improvement.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain common-cause and special-cause variation.
2. Select appropriate variable and attribute control charts.
3. Calculate and interpret control limits and process capability indices.
4. Distinguish statistical control from specification conformance.
5. Design integrated control plans and reaction plans.
6. Establish standard work, dashboards, audits, and layered process reviews.
7. Evaluate sustainment risks and design ownership, training, and change-management mechanisms.
8. Critically assess the limitations of SPC in modern, data-rich production systems.

1. Variation, Stability, and Capability

1.1 Common and special causes

Common-cause variation is the inherent, relatively stable variation produced by the combined effect of routine inputs, equipment, methods, people, materials, and environment. Reducing it generally requires redesigning or improving the process system.

Special-cause variation is attributable to a specific, non-routine influence, such as a broken tool, incorrect material, sensor failure, power interruption, or newly introduced work method. Special causes should normally be identified and removed before attempting to recalculate process limits.

A process is statistically stable when its observed variation is consistent with a stable probability distribution over the period examined. Stability does not mean that every observation is identical, nor does it guarantee acceptable output. A stable process can repeatedly produce defective products if its mean is poorly centred or its natural variation is too large.

1.2 Specifications and control limits

Specifications are externally or internally defined requirements, commonly expressed as a lower specification limit (LSL) and upper specification limit (USL). Control limits are statistically estimated boundaries describing expected process behaviour. Specifications answer, “Is the product acceptable?” Control limits answer, “Is the process behaving as expected?”

Confusing these concepts leads to two frequent errors: adjusting a stable process merely because an observation is near a specification limit, and assuming that a process is capable because no observation has yet exceeded a specification.

For an approximately normal individual measurement X, the natural process spread is often represented by μ ± 3σ, where μ is the process mean and σ is the within-process standard deviation. The interval contains approximately 99.73% of observations if normality and stable parameters are reasonable assumptions. The 3σ convention is a practical decision rule, not a universal law of nature.

1.3 Capability indices

For a stable process with two-sided specifications, potential capability is commonly measured by:

where USL and LSL are specification limits and σ is the estimated within-process standard deviation. Cp measures the ratio of allowable specification width to natural process width; it does not account for process centring.

Actual capability, including centring, is measured by:

Here μ is the process mean. These indices assume a stable process, meaningful specifications, an appropriate estimate of within-process variation, and—when interpreted using normal-distribution probabilities—an approximately normal distribution. For non-normal data, transformation, percentile-based methods, or a distribution-free capability analysis may be more appropriate.

For example, suppose USL = 10.20 mm, LSL = 9.80 mm, μ = 10.05 mm, and σ = 0.05 mm. Then:

The upper-side index is 0.15/(0.15) = 1.00, and the lower-side index is 0.25/0.15 = 1.67. Therefore Cpk = 1.00. The process has adequate potential spread relative to specifications, but its centring toward the upper limit reduces actual capability. The appropriate response is not indiscriminate variation reduction alone; centring and the economic consequences of the upper-tail risk must be addressed.

2. Control Chart Architecture

A control chart contains observations in time order, a centre line, and statistically derived upper and lower control limits. Time order is essential because randomised or aggregated data can conceal trends, cycles, and shifts.

A general control-limit form is:

where CL is the centre line, σ̂_s is the estimated standard deviation of the chart statistic, and k is commonly 3. The limits must be derived for the statistic being charted; they are not automatically equal to ±3 times the raw-data standard deviation.

Interpretation should use both point signals and non-random patterns. A point beyond a limit is strong evidence of a special cause under the chart’s assumptions. Runs, trends, cycles, stratification, and systematic alternation may also indicate non-random behaviour. Rules such as “eight successive points on one side of the centre line” are practical screening rules, but organisations should define them in advance and avoid indiscriminate use of many overlapping rules, which increases false alarms.

2.1 Variable charts

Variable charts are used for measured characteristics such as length, mass, temperature, torque, pressure, or cycle time. They preserve more information than pass/fail classification and are usually preferred when reliable measurement data are available.

#### X̄–R charts

The X̄–R chart is appropriate for rational subgroups, usually small samples collected under similar conditions. The X̄ chart monitors subgroup means and the R chart monitors within-subgroup ranges.

For subgroup size n, subgroup mean x̄_i, and subgroup range R_i:

The average range is R̄. Standard constants d₂, A₂, D₃, and D₄ depend on subgroup size and are obtained from standard SPC tables:

The constants assume independent observations within and between rational subgroups and are designed to estimate short-term variation. A negative calculated lower limit is conventionally set to zero for the range chart because a range cannot be negative.

#### X̄–S charts

For larger subgroups, the X̄–S chart uses subgroup standard deviations rather than ranges. It generally uses information more efficiently when subgroup sizes are moderate or large. The exact constants depend on the chosen standard-deviation estimator and subgroup size; therefore, software settings and reference tables must be verified rather than treated as interchangeable.

#### Individuals and Moving Range charts

The I–MR chart is suitable when measurements occur one at a time or rational subgrouping is not feasible. The moving range for consecutive observations is:

A common estimate of short-term standard deviation is:

For a moving range of two observations, d₂ is approximately 1.128. The individuals limits are then commonly calculated as:

This approach assumes observations are appropriately ordered and that the moving range reflects short-term variation. Strong autocorrelation, seasonality, or changing operating conditions can make interpretation misleading.

2.2 Attribute charts

Attribute charts classify or count outcomes. They are useful when measurement is costly, the characteristic is inherently categorical, or inspection produces counts such as defects, defective units, complaints, or failures.

For p-chart subgroups, p_i = d_i/n_i, where d_i is the number of defective units and n_i is the inspected sample size. If a stable baseline proportion p̄ is available:

Limits are truncated at 0 and 1 where necessary. For constant n, the np chart uses:

For a c chart, if c̄ is the average count per constant inspection unit:

For a u chart with ū average defects per unit and n_i inspected units:

The Poisson assumptions behind c and u charts require that defects are counts, opportunities are appropriately defined, and events are sufficiently independent. Overdispersion, clustering, or changing inspection sensitivity may require alternative modelling or process investigation.

2.3 Chart-selection logic

Selection should begin with the quality question, not with software availability. Ask whether the response is measured or classified, whether the observation is a unit or a defect, whether sample size varies, and whether rational subgroups exist. Also examine the data-generation process: a p chart for defective units is not equivalent to a u chart for defects because one unit may contain multiple defects.

A chart should have a clear operational owner, defined sampling frequency, traceable data source, validated measurement system, and pre-agreed reaction criteria. A sophisticated chart with unreliable data is less valuable than a simple chart based on trustworthy information.

3. Measurement and Data Integrity

SPC conclusions are only as credible as the measurement system. A variable measurement system should be evaluated for repeatability, reproducibility, bias, stability, linearity, and resolution. Attribute inspection systems require assessment of agreement within inspectors, between inspectors, and with a reference standard.

Data integrity includes correct timestamps, equipment identifiers, operator or shift information, lot traceability, units, and status of reworked or screened material. Automated collection reduces transcription errors but does not eliminate incorrect sensor calibration, software mapping errors, or intentional data manipulation. Access controls, audit trails, validation, and segregation of duties are therefore governance requirements as well as technical controls.

4. Control Plans and Reaction Plans

A control plan translates process knowledge into routine prevention and detection. A useful control plan identifies each process step, key input and output characteristics, specification or target, measurement method, sample size, frequency, responsible role, control method, and reaction requirement.

A reaction plan must be specific enough to prevent improvisation. It should answer five questions: What signal occurred? What product is suspect? Who stops, contains, or authorises continuation? What investigation is required? What evidence permits restart?

A robust reaction sequence is:

1. Protect the customer and downstream process.
2. Identify and segregate suspect material using traceability.
3. Verify the signal and check the measurement system.
4. Search for the assignable cause using process knowledge and data stratification.
5. Correct the cause and verify the process.
6. Disposition affected material through authorised procedures.
7. Document the event and update risk controls if necessary.

Control limits should not be recalculated immediately after every out-of-control point. Limits are estimates of baseline behaviour; changing them without removing the cause can conceal deterioration. Limit revision should follow documented analysis and approval.

5. Standard Work and Visual Management

Standard work defines the best currently known method for performing a task safely, consistently, and efficiently. It normally specifies takt or cycle expectations, work sequence, standard in-process stock, critical quality points, safety controls, and abnormality responses. It is not intended to suppress improvement. Rather, it creates a stable baseline from which improvement can be measured.

Effective standard work is visible at the point of use, written in language understood by operators, linked to actual equipment and revision status, and verified through observation. A document that is technically correct but unavailable, unreadable, or inconsistent with actual practice is not effective control.

Visual management should make normal and abnormal conditions apparent without extensive interpretation. Examples include status boards, reaction-limit indicators, first-piece approval markers, escalation lights, and digital displays. Visual systems must not become decorative dashboards; each displayed measure should have a decision or response associated with it.

6. Dashboards, Audits, and Layered Process Reviews

A production dashboard should combine outcome, process, and response measures. Outcome measures include defects per unit, first-pass yield, customer complaints, and capability. Process measures include temperature, torque, preventive-maintenance completion, changeover adherence, and error-proofing status. Response measures include time to contain, overdue corrective actions, and recurrence rate.

A useful dashboard follows the principle of one metric, one definition, one owner, and one action threshold. Excessive metrics create attention dilution and encourage local optimisation. Targets should not reward under-reporting defects or avoiding escalation.

Audits evaluate whether the management system is being followed and whether it remains effective. They should examine evidence, not merely signatures. A sound audit programme combines process audits, product audits, system audits, and periodic review of whether controls still address current risks.

Layered Process Audits (LPAs) use different organisational levels to examine a small number of critical standards at different frequencies. Operators and supervisors may conduct frequent checks; department managers and senior leaders conduct less frequent checks focused on systemic adherence. LPAs are most effective when they are brief, risk-based, respectful, and linked to rapid closure of findings. They should not be used as punitive surveillance or as a substitute for engineering problem solving.

7. Sustaining Improvements as a Socio-Technical System

Improvement decay often occurs because the technical solution is implemented without changing the surrounding system. Sustainment requires alignment among process design, leadership behaviour, skills, incentives, information, and accountability.

Ownership and governance

Every critical control requires a named process owner with authority, resources, and performance visibility. The owner maintains the control plan, approves changes, reviews signals, and ensures that corrective actions are effective. Governance should define escalation paths from operator to supervisor, engineering, quality, and executive levels.

Training and competence

Training should distinguish awareness, knowledge, demonstrated skill, and sustained competence. A practical cycle is explain, demonstrate, practise, observe, assess, and re-verify. Requalification is appropriate after major process changes, extended absence, repeated errors, or evidence that the standard is not being followed. Training records should identify the specific standard and version, not merely the topic title.

Change management

A process change can alter the mean, variation, measurement system, failure modes, or control limits. Change control should therefore include risk assessment, trial planning, validation, documentation revision, training, approval, and post-change monitoring. The Plan–Do–Study–Act cycle supports controlled learning: plan the change and measures, implement on a defined scale, study both intended and unintended effects, and standardise or revise the approach.

Sustaining through daily management

Daily management integrates tier meetings, visual performance review, abnormality escalation, and action tracking. The purpose is not to hold meetings for their own sake but to shorten the interval between signal, response, and learning. Leaders should routinely ask whether standards are clear, whether abnormal conditions are visible, and whether recurring problems are receiving systemic countermeasures.

8. Illustrative Case: Injection-Moulded Housing Line

The following is a fictional illustrative case.

Background and problem

A manufacturer of injection-moulded electrical housings experienced increasing customer complaints about dimensional fit. The critical width specification was 49.80–50.20 mm. Operators recorded one measurement every hour, but the plant used an undifferentiated “in-specification percentage” and had no formal reaction plan.

Analysis

A measurement-system study showed acceptable repeatability and reproducibility after recalibration of one gauge. An I–MR chart revealed a gradual upward shift after mould-cleaning events. Stratification by mould cavity showed that cavity 3 had higher variation. The process was not stable, so capability indices were not used as the sole basis for decisions.

Engineering inspection found partial blockage in a cooling channel. The blockage increased local temperature and altered shrinkage. The condition was intermittent because it depended on production duration and cooling-water quality.

Approach and solution

The team introduced a rational subgroup of five consecutive parts every two hours for each cavity and used X̄–R charts. The control plan specified cooling-water checks, mould-cleaning frequency, cavity identification, and gauge verification. The reaction plan required stopping the affected cavity, segregating material since the last confirmed good check, checking cooling flow, and obtaining authorised restart approval.

Standard work was revised with photographs of the correct cooling-channel inspection. A supervisor reviewed the chart at each shift handover. Maintenance owned the cooling-flow check, quality owned chart-limit governance, and production owned immediate containment. An LPA checked whether the reaction plan was understood and followed.

After corrective action, the process remained stable for eight weeks. The mean was centred near 50.00 mm and within-process variation decreased. Capability was calculated only after stability was demonstrated. The principal outcome was not merely a higher capability index; it was a controlled process with traceability, rapid response, and reduced recurrence.

Learning points

The case illustrates that an attribute percentage can conceal shifts and cavity-specific behaviour. Measurement-system verification must precede interpretation. Control charts, maintenance controls, standard work, ownership, and audits must operate as one system. Finally, sustainment depends on making the desired response easier and more reliable than informal workarounds.

9. Limitations and Contemporary Developments

SPC assumes that the data-generating process and measurement system are sufficiently understood. Non-independent observations, short production runs, product mix changes, rare events, autocorrelation, non-normal distributions, and changing inspection practices can invalidate conventional interpretations. In such settings, time-series methods, stratified charts, rare-event charts, survival analysis, or model-based monitoring may be more suitable.

Automated SPC and machine learning can detect subtle patterns across equipment, recipes, images, and sensor streams. However, predictive alarms require validation, explainability appropriate to the decision, cybersecurity, and controls against model drift. An algorithm that predicts an abnormality but cannot identify the affected product or trigger a feasible response may create noise rather than improvement. Human accountability remains essential for release, containment, and safety decisions.

Ethically, organisations must avoid manipulating sampling, suppressing inconvenient signals, or using charts primarily to blame individuals for system variation. Quality records may have regulatory significance, particularly in aerospace, medical devices, automotive, food, and pharmaceutical production. Data retention, electronic signatures, access control, and audit trails should comply with applicable standards and law.

Best-Practice Checklist

• Validate the measurement system before trusting the chart.
• Define the unit, opportunity, subgroup, and sampling rationale.
• Use time-ordered data and preserve traceability.
• Separate stability analysis from specification capability.
• Establish reaction plans before signals occur.
• Do not recalculate limits to hide special causes.
• Link every critical control to an owner and escalation route.
• Keep standard work current, visible, and observable.
• Use dashboards to drive decisions, not merely display results.
• Conduct risk-based audits and LPAs without creating fear.
• Train, assess, and requalify people after meaningful changes.
• Monitor sustainment and recurrence, not only initial project gains.

Chapter Summary

SPC provides a structured way to understand process variation and distinguish routine behaviour from special causes. Variable charts such as X̄–R, X̄–S, and I–MR charts are selected according to measurement type and subgrouping; attribute charts such as p, np, c, and u charts are selected according to whether the data represent defective units or defects and whether opportunity varies. Control limits describe process behaviour, whereas specifications describe requirements. Capability indices are meaningful only when the process is stable and the data and assumptions are appropriate.

Effective improvement control extends beyond charts. Control plans define what must be controlled, reaction plans define what must happen when abnormality occurs, and standard work establishes the current best method. Dashboards, audits, and layered process reviews create visibility and accountability. Sustainment requires ownership, validated measurement, competence, disciplined change management, leadership routines, and ethical data governance. Technology can improve detection, but it cannot replace sound process knowledge and human responsibility.

Key Terms

Statistical Process Control: A method for monitoring and improving process behaviour using time-ordered data.

Common Cause: Routine variation inherent in the process system.

Special Cause: Non-routine variation attributable to a specific influence.

Control Limit: A statistically derived boundary used to identify unusual process behaviour.

Specification Limit: A customer, engineering, or regulatory acceptance boundary.

Rational Subgroup: A sample formed so that within-group variation represents short-term process variation.

Variable Data: Quantitative measured data on a continuous or discrete scale.

Attribute Data: Classified or counted data.

X̄–R Chart: A chart pair for subgroup means and ranges.

I–MR Chart: A chart pair for individual observations and moving ranges.

p Chart: A chart for the fraction or proportion defective.

c Chart: A chart for defect counts with constant opportunity.

u Chart: A chart for defects per unit when opportunity may vary.

Process Capability: The degree to which a stable process can meet specifications.

Cp: Potential capability based on specification width and process spread.

Cpk: Capability index incorporating process centring.

Control Plan: A structured description of process controls, measurements, owners, and responses.

Reaction Plan: Predefined actions for abnormal signals or conditions.

Layered Process Audit: A risk-based audit performed at multiple organisational levels.

Sustainment: Institutionalisation of improvements so that performance does not regress.

Review Questions

Short Answer Questions

1. Distinguish common-cause variation from special-cause variation.
2. Why must control limits not be confused with specification limits?
3. When is an I–MR chart appropriate?
4. What is the difference between a defective unit and a defect?
5. Name four elements that should appear in a reaction plan.

Long Answer Questions

1. Explain the selection, assumptions, and interpretation of X̄–R, I–MR, p, c, and u charts.
2. Discuss why capability analysis should normally follow stability analysis.
3. Design a control plan for a high-risk production operation.
4. Evaluate the role of standard work, dashboards, audits, and LPAs in sustainment.
5. Discuss the organisational and ethical conditions required for reliable SPC.

Analytical / Application Questions

1. A process has USL = 25.20, LSL = 24.80, μ = 25.10, and σ = 0.05. Calculate Cp and Cpk, interpret the difference, and recommend an action.
2. A p chart shows one point beyond its upper control limit after a machine change. Develop a reaction plan covering containment, verification, investigation, restart, and documentation.
3. A factory has stable overall defect performance but recurring failures on one shift and one product family. Explain how you would stratify the data, redesign the monitoring system, and sustain the corrective action.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a control limit from a specification limit?

A. A control limit is set by the customer, while a specification limit is calculated statistically.

B. A control limit describes expected process behaviour, while a specification limit defines an acceptance requirement.

C. Both limits have the same purpose but different names.

D. A specification limit can be recalculated after every production shift.

Correct Answer: B. A control limit describes expected process behaviour, while a specification limit defines an acceptance requirement.

Question 2

Which chart is generally appropriate for individual measurements collected one at a time?

A. np chart

B. c chart

C. I–MR chart

D. X̄–R chart with subgroups of ten

Correct Answer: C. I–MR chart

Question 3

A p chart is designed to monitor:

A. The proportion of defective units

B. The number of defects per constant area only

C. The mean of continuous measurements

D. The range within a rational subgroup

Correct Answer: A. The proportion of defective units

Question 4

If Cp is substantially greater than Cpk, the most likely interpretation is that:

A. The process is necessarily unstable.

B. The process has no specification limits.

C. The process has potential spread capability but is not well centred.

D. The measurement system is automatically invalid.

Correct Answer: C. The process has potential spread capability but is not well centred.

Question 5

For a c chart, the inspection opportunity should generally be:

A. Constant

B. Zero

C. Randomly omitted

D. Different for every observation

Correct Answer: A. Constant

Question 6

What should normally occur first when a control-chart point signals a special cause?

A. Delete the point from the dataset

B. Immediately widen the control limits

C. Contain potentially affected output and verify the signal

D. Change the specification limits

Correct Answer: C. Contain potentially affected output and verify the signal

Question 7

Which feature is essential to a rational subgroup?

A. It combines unrelated products to increase sample size.

B. It represents conditions intended to estimate short-term variation.

C. It is always collected once per month.

D. It excludes consecutive observations.

Correct Answer: B. It represents conditions intended to estimate short-term variation.

Question 8

The primary purpose of standard work is to:

A. Prevent all future improvement activity

B. Define the current best known method and create a stable baseline

C. Replace measurement-system validation

D. Transfer all accountability to the quality department

Correct Answer: B. Define the current best known method and create a stable baseline

Question 9

Layered Process Audits are most effective when they:

A. Are punitive and focused on finding individual blame

B. Use identical frequency and scope at every management level

C. Check critical standards at different organisational levels and enable rapid correction

D. Replace engineering problem solving

Correct Answer: C. Check critical standards at different organisational levels and enable rapid correction

Question 10

Why should capability indices generally be calculated only after stability is established?

A. Capability indices require no specifications.

B. An unstable process does not have a reliable single estimate of its routine behaviour.

C. Control charts cannot be used with variable data.

D. Capability indices apply only to attribute data.

Correct Answer: B. An unstable process does not have a reliable single estimate of its routine behaviour.