SPC Has Been Building Manufacturing Intelligence for Decades

By Harri Pihlajavesi, CEO | Predisys

In the late 1990s, I worked with dimensional measurement and Statistical Process Control in an automotive body shop. At one station, an inspector took gap-and-flush measurements from an unpainted steel vehicle body using a handheld gauge connected to a computer strapped to her wrist.

The measurements were transferred through a docking station and became part of the statistical understanding of the process, connecting what was physically happening on the factory floor with data, variation and process knowledge.

The technology has changed considerably since then. The underlying problem is familiar.

We want to understand how a manufacturing process is behaving, recognize when that behavior changes and have enough information around the change to determine what is happening.

SPC has been helping manufacturers do this for decades, and much of what we now describe as Manufacturing Intelligence has been developing alongside it.

Understanding process behavior

Every manufacturing process has variation.

If we measure a characteristic repeatedly, the values will not be identical. The important question is whether the variation we are seeing is consistent with the way the process normally behaves or whether something has changed.

This is one of the fundamental ideas behind SPC.

A control chart gives us a way to look at measurements over time and identify statistical evidence that the process has changed. The signal may be a point beyond a control limit, a shift in the process average, a trend or another pattern that would be unlikely to occur as part of the process's normal variation.

This gives us information that specifications alone cannot provide.

A specification tells us whether a product characteristic meets a requirement. Process behavior tells us something about the process producing that characteristic.

A process can begin changing while every measurement remains within specification. If we recognize the change early enough, we have an opportunity to investigate it before it results in scrap, rework, yield loss or a customer problem.

That has always been one of the practical reasons for doing SPC well.

The formulas are the easy part

The mathematics behind control charts has been understood for a long time. The more difficult part of implementing SPC across a real manufacturing environment is knowing what analysis applies to what data.

A measurement needs context.

Which product is being produced?

Which characteristic are we measuring?

Which process produced it?

Which machine, line, tool or cavity was involved?

Which specification applies?

Which control limits should be used?

Is this the beginning of a new production run or a continuation of the previous one?

What happened during previous runs of the same product and process?

In a simple example, we can draw a control chart from a column of numbers. In production, those numbers belong to something.

That structure becomes particularly important as manufacturing gets more complex. One plant may produce hundreds or thousands of products across many processes, machines, tools and cavities. Specifications may vary by product or customer. The same characteristic may be produced on several pieces of equipment, each with its own process history.

A real-time SPC system has to maintain those relationships continuously as data arrives. Over time, that creates a structured statistical history of manufacturing.

Control limits carry process knowledge

Control limits are a good example of how knowledge begins to accumulate around a process.

They are based on process behavior, so a useful set of limits reflects something the organization has learned from previous production.

The same is true of many of the decisions surrounding an SPC implementation.

Which characteristics are important enough to monitor?

How should the data be grouped?

When should limits be recalculated?

Who is allowed to change them?

What constitutes a meaningful signal for this process?

What should happen when one occurs?

These decisions combine statistical principles with knowledge of the manufacturing process.

As the system operates, the history becomes richer. We can see how a process behaved during previous runs, when changes occurred, whether the same pattern has appeared before and how stable the process has been over time.

This becomes useful when something changes because we have a basis for comparison. If a process begins shifting today, we are not looking at today's measurements in isolation. We can compare them with what we already know about that process.

A statistical signal starts an investigation

Knowing that a process changed does not tell us why it changed.

Suppose a control chart shows a shift in a dimensional characteristic at 10:42 on Tuesday morning.

Now the manufacturing questions begin:

What happened around 10:42?

Did the material lot change?

Was an adjustment made?

Did maintenance work on the equipment?

Was a tool replaced?

Did one process parameter move at approximately the same time?

Is the shift present across the machine or isolated to one cavity?

Are other characteristics showing related behavior?

The statistical signal gives us a place to start looking. The answer usually comes from connecting that signal to the physical and operational context surrounding the process.

This is where SPC naturally begins to connect with a broader idea of Manufacturing Intelligence. We have statistical knowledge about how the process behaves, manufacturing context that tells us what produced the data and operational information about what was happening when the behavior changed.

What happens after a signal matters too

In a working manufacturing environment, somebody usually needs to respond when a meaningful process change is detected.

An operator may acknowledge an alarm. A process engineer may investigate. Maintenance may inspect the equipment. Production may continue under observation or be stopped depending on the situation. A quality issue may be opened, or a supplier may need to become involved.

An action is eventually taken. Then we continue collecting data.

The process behavior after the action tells us something important. If the process returns to its previous behavior, that supports what we learned during the investigation. If nothing changes, we need to keep looking. If the problem disappears temporarily and later returns, that becomes part of the history too.

Capturing this sequence gives us more than the original statistical signal. We have the process behavior that led to the investigation, what people found, what they did and what happened afterward.

The next time a similar pattern appears, that history can be useful.

SPC can connect information across manufacturing

Another advantage of having consistent statistical information is that we can begin looking beyond a single control chart.

Suppose the same component is produced on several machines.

One machine begins behaving differently from the others.

Or several characteristics on the same tool shift at approximately the same time.

Perhaps the same pattern appears at two plants using material from the same supplier.

Those relationships can be difficult to recognize when each chart is treated independently. They become easier to investigate when the statistical information retains its manufacturing context.

This also changes the scale at which SPC can be useful.

A process engineer may be interested in a particular characteristic on one machine. A factory manager may want to know which processes are showing unusual behavior today. A quality leader may be looking across plants, products or suppliers for patterns that deserve attention.

The statistical foundation is related, while the question and level of detail change depending on who is looking at it.

Real-time SPC has been creating a history of manufacturing

When I think back to that automotive body shop in the 1990s, the gap-and-flush measurements were valuable because they became part of something larger than an individual inspection.

They contributed to an understanding of how the body-building process was behaving.

Modern manufacturing gives us far more information to work with.

Measurements can arrive directly from machines, gauges, test equipment, vision systems and other automated sources. We can collect process parameters alongside quality characteristics and connect them with production, equipment, material and supplier information.

The same basic question remains: what is the process telling us?

A mature SPC environment can provide years of information about process behavior. It can tell us what normal looked like, when meaningful changes occurred and how those changes related to particular products, equipment, tools, cavities or production runs.

When investigations and actions are connected to that history, we also begin retaining more of what people learned when they responded to those changes.

That accumulated understanding is an important part of Manufacturing Intelligence.

What this means for Manufacturing AI

AI creates an interesting new way to work with the process knowledge manufacturers have already been building.

Consider the difference between asking an AI model a general question about SPC and asking a question about a specific production process.

Why did this characteristic shift during yesterday's run?

Has this pattern occurred before?

Which other parameters changed around the same time?

Is the behavior isolated to this machine?

What happened the last time we saw it?

Which processes are showing statistically unusual behavior right now?

Answering those questions requires the AI to work with the statistical history of the actual manufacturing processes involved, along with the context that tells it what the data represents.

It also requires some care.

A statistical correlation between two changes does not establish that one caused the other. An AI agent should be able to help an engineer find relevant relationships and previous experience without turning every correlation into an explanation.

The statistical foundation matters here because it gives the AI something more useful than a collection of raw measurements. There is already information about expected process behavior, meaningful change and manufacturing context.

SPC has been developing that understanding one production run at a time for decades.

As Manufacturing AI develops, we have an opportunity to make much more of that history available when engineers and manufacturing leaders are trying to understand what is happening now.