SPC in Injection Moulding: Control Charts That Catch Drift Early

Why a good part is not the same as a controlled process

A part inside tolerance proves that the process was inside tolerance at the moment that part was made. It says nothing about the next part. A process running at the top of the tolerance band with a drift of two microns per hour will produce good parts for a shift and out of tolerance parts by the next one, and every inspection in between will pass.

Statistical process control addresses exactly that gap. It does not inspect the part; it inspects the process, by plotting a measurement over time against limits derived from the process's own variation. The question it answers is not whether this part is good but whether the process that made it is still the process that was qualified.

The distinction matters because the two questions have different answers and different costs. Inspection is a filter at the end of the line and it catches defects one at a time. A chart is a signal at the point of production and it catches a change once. A plant that only inspects will find the drift after the parts are made; a plant that charts will find it while there are still ten parts to correct.

There is a precondition that is usually skipped. A control chart assumes the measurement system is capable. If two operators measuring the same part differ by more than a third of the tolerance, the chart is plotting gauge variation rather than process variation, and every signal it produces is noise. Do the gauge study first, then chart.

Which variable to chart

Chart the variable closest to what the customer cares about, measured as directly as possible. For a dimensional part that is a critical dimension. For a sealed part it is often a weight, because weight integrates every cavity condition into one number and it can be measured automatically. For a part where appearance is the requirement, chart the process parameter that drives appearance rather than the appearance itself.

Weight deserves its reputation. Shot weight responds to melt viscosity, cavity pressure, hold time and mould temperature all at once, and it correlates with dimensions on almost every part. An in-line weight check on a robot or a conveyor gives a hundred percent sampling of a variable that a dimensional check can only sample. The chart of shot weight will show a drift days before a dimension leaves tolerance.

Chart the variable, not the average of the variable, unless the average is what the customer receives. Averaging hides the variation within a shot, and in a multi-cavity tool the difference between cavities is often larger than the drift over time. In that case chart each cavity separately, or chart the difference between the highest and lowest cavity, which is the number that decides whether the part is acceptable.

Avoid charting a parameter that the operator can adjust during the run without recording it. If hold pressure is tweaked to fix a short shot and the chart continues as if nothing happened, the limits no longer describe the process. Either lock the parameter or annotate the chart at the moment of the change.

Setting the limits before you have data

There is a chicken and egg problem at the start: limits should come from the process, but a new process has no history. The practical answer is to run a short capability study first. Thirty to fifty consecutive parts, measured in the order they were made, gives a standard deviation good enough to set provisional limits, which are then replaced by the real ones after a few weeks of production.

The provisional limits should be calculated, not chosen. Take the mean of the study, add and subtract three standard deviations, and use those as the control limits. Do not use the drawing tolerance as a control limit. Tolerance is what the customer accepts; control limits are what the process does. A process that is capable has control limits well inside the tolerance, and the gap between them is the room you have to react.

Plot the tolerance on the same chart, as two separate lines outside the control limits. The visual gap between them is the most useful thing on the page: it tells the operator how much drift is left before the part is actually bad. A chart where the control limits touch the tolerance lines is a chart of a process with no margin, and the correct response is to improve the process rather than to tighten the reaction.

Recalculate the limits when the process changes, and only then. A new mould, a new material, a new machine or a deliberate change to the cycle is a reason to rebuild the baseline. Recalculating because a point fell outside is not: that removes the evidence that the process moved, which is the one thing the chart exists to provide.

Table 1 — Five charts cover almost every moulding case

The rules that turn points into a signal

A single point outside three sigma is a signal, and it is the only rule most plants use. That is a mistake, because the most expensive drift is the one that stays inside the limits. A process that has moved by one sigma produces almost no points outside three sigma, and yet every part is being made off-centre.

The run rule catches that case. Eight or nine consecutive points on the same side of the centre line is a signal, even though every point is inside the limits. The probability of that happening by chance is about one in a few hundred, so it is far more likely to be a real shift. The chart in this article shows exactly that pattern: nine samples above the mean, then the curve breaks through the upper limit.

The trend rule catches a slower movement. Six or seven points that consistently increase or decrease is a signal, regardless of where they sit relative to the limits. A tool that is warming up, a material that is absorbing moisture, a check ring that is wearing: all of these produce a trend long before they produce a violation.

Two out of three consecutive points beyond two sigma, on the same side, is a fourth rule worth using. It is more sensitive than the run rule and less noisy than a single point rule, and it is a reasonable compromise for a process with moderate variation. Whatever set of rules is chosen, write them on the chart so the operator applies them the same way every shift.

Table 2 — Four rules and one warning

Reacting to a signal without over-reacting

The commonest failure of an SPC programme is not a missed signal. It is a reaction to every signal. An operator who adjusts the machine each time a point wanders produces a process that is worse than the one that was left alone, because the adjustments themselves become the largest source of variation. This is over-control, and it is visible on a chart as a saw-tooth that follows the points.

The rule is to react to signals, not to points. A point inside the limits, even a point far from the mean, is not a reason to touch the machine. A run, a trend or a point outside the limit is. Between those two states there is nothing to do, and doing nothing is a skill that has to be taught.

When a signal does occur, the response should be a sequence rather than an improvisation. Confirm the measurement first, because a gauge that has drifted produces the same pattern as a process that has drifted. Then check the obvious process inputs in order: material lot, mould temperature, hold pressure, back pressure, cycle time. Then look for a special cause in the tool or the machine. Only after all of that should the settings be changed.

Record the reaction on the chart. A chart with a note saying what was changed and when is a diagnostic record; a chart with a break in the line and no explanation is a puzzle. Six months later, when the same pattern appears, the note is the difference between solving it in an hour and solving it in a week.

What SPC does not fix

SPC detects change; it does not create capability. A process that is stable at a Cpk of 0.8 will produce a beautiful, in-control chart and an unacceptable defect rate. The chart's job in that case is to prove the process is stable so that the improvement effort goes into the process rather than into hunting special causes that do not exist.

It also does not replace inspection where inspection is required. Regulated products, safety-critical dimensions and customer-mandated sampling all need the records that inspection produces. SPC reduces the amount of inspection needed by making the process predictable, but it does not remove the obligation.

And it does not survive without discipline. A chart that is filled in at the end of the shift from a pile of parts, rather than as the parts are made, has no time information in it and cannot detect a drift. The value of the chart is entirely in the order of the points, and a chart filled in retrospectively is a work of fiction.

The realistic goal is a chart that is looked at by the operator during the run, not by a manager at the end of the month. That means a chart at the machine, in the operator's line of sight, with limits that are current and a pen that works. Everything else is documentation.