Manufacturing Process Control: SPC, APC & Real-Time Process Data

What is process control?

Process control is the discipline of keeping a manufacturing process operating inside a defined envelope of parameters — temperature, pressure, speed, torque, cycle time, dimensional tolerances — so that the output meets specification without post-hoc inspection. It is the difference between producing a part and then measuring whether it is good, versus monitoring the process that produces the part so that you know it will be good before it leaves the machine.

Before going further, a terminology note that accounts for most of the confusion around this topic in English-language searches. "Process controlling" is a German loan translation (Prozesscontrolling) that typically points at business-process monitoring in the controlling/finance sense — KPI dashboards for end-to-end administrative workflows. "Process control", without the -ing, is the canonical English engineering term for shopfloor parameter control. These are different disciplines with different tools, different audiences, and different failure modes. This article is about the second one, because that is what actually matters in a factory. The first is covered adequately by any generic BPM vendor.

Process control vs. process controlling — the terminology trap

The table below separates the two meanings explicitly. If you landed here looking for the first one, the MES-and-analytics world is not your world; look for Business Process Management (BPM) tooling instead.

Quality control is the one most often conflated with process control, and the distinction is worth hammering. Quality control measures the part after it has been made and decides whether to ship it. Process control measures the process while the part is being made and decides whether the machine is still operating inside specification. A plant that relies only on quality control will always know what it scrapped yesterday; a plant that practises real process control will know what it is about to scrap in the next ten minutes — and can intervene before the scrap happens.

The three layers of manufacturing process control

In practice, "process control" is not one thing but three nested capabilities, each with a different purpose and a different data requirement. Conflating them is why most discussions of the topic go in circles.

BPC is where the machine lives. It is the PID controller that holds an injection-moulding barrel at 220 °C, the interlock that stops a press if the light curtain is broken, the deterministic ladder logic that sequences a weld cycle. BPC is a prerequisite — if BPC fails, nothing else matters — but it is invisible from the MES layer. It runs in the PLC, it runs fast, and it has no opinions about statistics.

APC sits above BPC. It takes a model of the process (often an empirical model fitted from historical data, sometimes a first-principles model from chemistry or physics) and uses it to choose setpoints that BPC then enforces. In batch chemistry it shows up as Model Predictive Control adjusting temperature trajectories; in discrete manufacturing it shows up as adaptive feed-rate control in CNC machining, or closed-loop cycle-time optimisation in injection moulding.

SPC is what most manufacturing engineers mean when they say "process control" without qualification. It is the Shewhart tradition: sample the process, plot the measurement on a control chart, apply the Western Electric rules to decide whether the process has shifted, and intervene if it has. SPC does not actively control the machine; it tells a human when the process has drifted far enough that the human should control the machine. That is less automated than APC, but it is where the Cpk and Ppk numbers that customers demand actually come from.

Where the data has to come from — and why it usually doesn't

All three layers need data, and the quality of each layer is bounded by the quality of the data feeding it. This is where most "process control" implementations quietly fail. A typical discrete-manufacturing plant has plenty of process data in the PLC — cycle times, motor currents, barrel temperatures, clamp forces, setpoints, feedbacks — and none of it leaves the PLC. It is used for BPC, discarded, and never seen by anyone upstream.

For SPC to work, that data has to be continuously lifted out of the PLC, timestamped, normalised, and streamed to a place where statistical analysis can happen. In the architecture I designed for the SYMESTIC platform, this is a three-stage path: OPC UA or digital-I/O capture at the machine, an edge gateway that timestamps and buffers, and a cloud data layer that normalises and persists the stream. A few design points that matter:

• Sample rate selection. A 10 Hz sample on a process variable that actually changes at 1 Hz is waste; a 1 Hz sample on a variable that changes at 20 Hz is blindness. Getting this right per signal is the difference between useful SPC charts and charts that miss every transient.
• Deadbanding and compression. Not every sample needs to make the round trip to the cloud. Well-designed deadbands reduce payload by 80��95 % without losing any statistical signal — but only if the deadband is configured per signal class, not globally.
• Edge buffering for network partitions. A factory network will drop. An edge gateway that buffers locally and catches up when the link returns is the difference between an SPC chart with gaps and an SPC chart with integrity.
• Clock discipline. Cross-correlating a temperature trace from one PLC with a force trace from another requires synchronised clocks to millisecond precision. In practice this means NTP or PTP across the plant network, and an edge that refuses to timestamp if its clock is unsynchronised.

These are not academic points. In the Neoperl implementation, the measurable 15 % quality-loss reduction came specifically from correlating PLC alarm traces with process-parameter traces across shared time axes — which was only possible because the gateway architecture enforced synchronised timestamping at source. Without that, the correlation would have been arithmetic noise.

Reactive vs. predictive process control

Here is the thesis that separates modern process control from the textbook version: most "process control" in the field is still reactive, and the gap between reactive and predictive is almost entirely a data-infrastructure problem, not an algorithm problem.

Reactive process control says: the SPC chart shows a point outside the upper control limit, so the process has shifted, so we should investigate. That is a generation ahead of pure quality control — which would wait until the finished part failed dimensional inspection — but it still reacts to a symptom after the process has already drifted.

Predictive process control says: the bearing temperature on axis 3 is 4 °C above nominal and trending upward at 0.2 °C per cycle, historically this pattern precedes a dimensional drift by 15–20 cycles, therefore intervene now. The algorithms for this are not exotic — Western Electric rules extended with trending rules, or simple rolling-window regression — but they are useless without continuous, high-fidelity process data. The capability is bounded by the data layer, not by the statistics.

Plants that invest in process control without first fixing their process-data capture almost always end up with expensive SPC licences running on sparse, manually-entered samples — which delivers the reactive version at the cost of the predictive one. The correct sequence is: capture the process data automatically and continuously first, then layer SPC and APC on top. Reversing that sequence is the single most common failure mode I see in the field.

FAQ

What's the difference between process control and quality control?
Process control monitors the process parameters that produce the part — temperature, pressure, cycle time, machine variables — in order to keep the process inside specification. Quality control measures the finished part. The first prevents defects; the second catches them. Both matter, but a plant that relies only on quality control pays for defects it could have prevented.

Is SPC the same as process control?
No. SPC (Statistical Process Control) is one layer of process control — the statistical monitoring layer. Process control in its full sense also includes Basic Process Control (PLC-level closed loops running in milliseconds) and Advanced Process Control (model-based setpoint optimisation running in seconds). SPC is the most visible layer because it produces the charts managers look at, but it is not the whole discipline.

Do I need an MES for process control?
Not strictly — SPC has existed since the 1920s, long before MES existed as a category. But an MES is how modern process control becomes continuous and automated rather than sampled and manual. Without an MES-class data layer, SPC in a plant with more than a handful of machines collapses into paper forms and Excel — which is process control in name only.

How much process data is actually needed?
Less than most vendors suggest, more than most plants currently capture. For discrete-manufacturing SPC, 1–10 Hz per relevant process variable is usually sufficient, with 10–50 variables per machine being typical. That is 10–500 samples per machine per second — trivial for modern cloud-MES architectures, but often impossible for the Excel-based processes it replaces.

Can process control prevent defects entirely?
No manufacturing process reaches zero defects, but well-instrumented process control routinely reduces scrap by 30–60 % versus inspection-only approaches. The mechanism is simple: most defects have a process-parameter signature that precedes the defect by minutes to hours. Catching that signature and intervening is cheaper than scrapping the part and investigating afterwards.

What's the right starting point?
Continuous automated capture of process data from the machines you already have. Without that data layer, SPC and APC are paper exercises. With it, layering SPC on top is a matter of configuration rather than implementation. SYMESTIC's Process Data and Alarms modules exist specifically for this starting point.

Related: Process Quality · Production Defect · Production Efficiency · OEE · MES · SYMESTIC Process Data · SYMESTIC Alarms

About the author

Mark Kobbert

CTO of SYMESTIC GmbH since 2020. B.Sc. Wirtschaftsinformatik. Responsible for the cloud-native MES platform architecture on Microsoft Azure — microservices, OPC UA and MQTT connectivity, edge gateways, and real-time process-data pipelines running at 15,000+ machines across 18 countries. Specialist in IT/OT convergence, brownfield machine connectivity without PLC intervention, and the data-layer design that makes continuous process control possible at scale. · LinkedIn