SPC

What is SPC?

SPC (Statistical Process Control) is a method that uses statistical techniques, above all control charts, to monitor the variation in a production process so that you can tell normal noise from a real disturbance and intervene before products fall outside specification. SPC therefore looks not only at the finished product but at how the process behaves over time. The method is more than a century old, yet it is enjoying a second life now that measurement data from PLCs, SCADA and the historian is available automatically.

🧠 How does SPC work?

Every process varies. SPC distinguishes two kinds of variation:

• Common causes - the natural, predictable noise of a stable process: small differences in raw material, temperature or wear. You do not react to these measurement by measurement; you only reduce them by improving the process itself.
• Special causes - an identifiable disturbance, such as a worn blade, a wrongly set machine or a new batch of raw material. These call for immediate investigation and action.

A process with only common causes is said to be in statistical control. In 1924 Walter Shewhart, then at Western Electric and later at Bell Telephone Laboratories, devised the control chart for this purpose: a time series of measurements with a centre line and upper and lower limits at three standard deviations (±3σ). For a normally distributed, stable process about 99.73% of the points then fall within the limits, so a point outside them is a strong signal that something has changed.

It is important to distinguish control limits from specification limits. Control limits are calculated from the process itself (the “voice of the process”); specification limits come from the customer or the design (the “voice of the customer”). A process can be in control and still produce scrap, or stay within specification while the control chart is already raising an alarm.

📊 Which control charts are there?

The choice depends on the type of data and the size of the sample (subgroup):

Shewhart charts are specified in ISO 7870-2, whose second edition was published in 2023. Alongside the control chart, histograms, Pareto analysis, scatter diagrams and the fishbone diagram belong to the classic quality tools used with SPC, among other things for root cause analysis.

🚨 When is a process out of control?

A single point beyond ±3σ is the best-known signal, but patterns say a lot too. The Western Electric rules from the Statistical Quality Control Handbook (1956) and the Nelson rules (1984) make that judgement objective. Commonly used signals:

• One point outside the control limits
• Two out of three consecutive points beyond 2σ on the same side
• Four out of five points beyond 1σ on the same side
• Eight points in a row on one side of the centre line (shifted mean)
• Six points in a row steadily rising or falling (a trend, for instance wear)

More rules mean earlier detection, but also more false alarms. With the 3σ rule alone you get on average one false alarm per roughly 370 points; adding the Western Electric rules brings that to about one per 92 points. Anyone who puts SPC alarms into an MES or SCADA system should therefore choose them as carefully as in alarm management.

📐 What do Cp and Cpk tell you?

Only once a process is in control does it make sense to ask whether it is also capable: does it produce within tolerance? Two indices answer that question, described in the ISO 22514 series:

• Cp = (USL − LSL) / 6σ - does the spread fit within the tolerance, regardless of where it sits?
• Cpk = min(USL − μ, μ − LSL) / 3σ - the same, but corrected for a process that is off-centre

In automotive (IATF 16949 with the AIAG core tools SPC and PPAP) an index above 1.67 in the initial process study is considered sufficient and between 1.33 and 1.67 possibly acceptable; customers may set stricter requirements. Pp and Ppk are the variants based on the total, long-term spread.

🔧 How do you introduce SPC, step by step?

1. Choose critical characteristics - start with characteristics that affect quality or safety, for example from an FMEA or a HACCP plan.
2. Secure the measurement system - use calibrated instruments and check with a measurement system analysis (gauge R&R) whether the measurement error is small enough relative to the tolerance.
3. Define rational subgroups - take items made in quick succession, so that variation within the subgroup contains only common causes.
4. Phase I: establish the limits - collect about 20 to 25 subgroups, remove explained outliers and calculate the control limits.
5. Phase II: monitor - plot new measurements against fixed limits and record cause and action for every signal.
6. Improve and recalculate - a deliberate process improvement calls for new limits; this is how SPC becomes part of PDCA and Six Sigma (DMAIC).

Worked example. A filling line produces 500 g packs; every fifteen minutes a checkweigher weighs five packs. The grand mean of the subgroup means is 503.0 g and the average range R̄ is 4.0 g. With the constant A₂ = 0.577 for n = 5 the limits are 503.0 ± 2.3 g, i.e. from 500.7 to 505.3 g. The process spread σ ≈ R̄/d₂ = 4.0/2.326 ≈ 1.7 g. Under the European e-mark rules (Directive 76/211/EEC) the average must be at least the nominal quantity, only a small share of packs (in practice no more than 2.5%) may be more than 15 g underweight, and no pack may be more than 30 g underweight. SPC shows how close to 500 g you can set the target weight without breaking those rules: every gram less overfill is a direct saving.

🏭 How does SPC fit into the IT/OT environment?

Operators used to draw control charts by hand. Today the data comes automatically from sensors, scales and vision systems via OPC UA or MQTT, is stored in a historian or time series database, and MES software or a dedicated SPC package calculates the charts in real time. Laboratory results come from the LIMS. In the ISA-95 model SPC belongs to quality operations management at level 3.

Two pitfalls deserve attention. First, SPC is not a control loop: a PID controller or closed-loop control corrects every deviation immediately, whereas SPC explicitly warns against adjusting for common-cause noise (tampering), which increases the spread. Second, automated SPC is a computerised system: in pharmaceuticals it falls under GMP, must be validated according to GAMP, and its electronic records must comply with 21 CFR Part 11 and the requirements for data integrity.

📈 What are the benefits of SPC?

• Faster problem detection - shifts become visible before scrap is produced
• Less scrap and rework - which feeds straight into the quality factor of OEE
• Less overfill and material loss - because you can steer closer to the target
• Data-driven improvement - decisions based on facts rather than gut feeling
• Broadly applicable - from discrete manufacturing in metalworking and automotive to food, pharmaceuticals and the process industries (chemicals, petrochemicals), where parameters are often monitored per batch; see also discrete, batch and continuous processes
• Better compliance - ISO 9001 requires monitoring and analysis of processes; the FDA process validation guidance of 2011 names continued process verification as its third stage and EU GMP Annex 15 calls for ongoing process verification, for which SPC is the standard monitoring instrument

❓ Frequently asked questions

What is the difference between SPC and end-of-line quality inspection?

End-of-line inspection checks the finished product and sorts out rejects after the problem has already occurred. SPC monitors the process during production and flags a shift before any scrap is made. SPC prevents defects, whereas final inspection only finds them.

Why are SPC control limits set at three sigma?

Three sigma is an economic choice made by Shewhart: for a stable, normally distributed process 99.73% of the points then fall within the limits. That gives you few false alarms, on average one per roughly 370 points, while still offering a good chance of spotting real disturbances.

How many measurements do you need to start with SPC?

A rule of thumb for establishing SPC control limits is 20 to 25 subgroups, for example 25 subgroups of five items. With less data you can use provisional limits and recalculate them once more measurements are available.

What is a good Cpk value?

A Cpk of 1.33 is regarded in many industries as the minimum for a capable process in ongoing production. Automotive customers often require 1.67 for new processes, and a Cpk of 2.0 corresponds to the Six Sigma level.

Is SPC the same as Six Sigma?

No, SPC is a tool, whereas Six Sigma is a complete improvement methodology with projects, roles and the DMAIC cycle. Six Sigma mainly uses SPC in the Control phase to keep an improved process stable.

Can machine learning replace SPC?

Machine learning complements SPC, especially with many correlated variables where classic control charts fall short. For individual critical characteristics SPC remains the standard, because the method is transparent, explainable and recognised by auditors and standards.

📌 In summary

SPC makes process variation visible with control charts, separates noise from real disturbances and uses Cp and Cpk to show whether a process can produce within tolerance. With automatic data from the OT layer and the MES it becomes real-time quality monitoring that reduces scrap, overfill and audit risk.

Statistical Process Control (SPC) uses statistics to check whether a process is stable and capable, so that you act on real disturbances rather than on noise.