Process drift is one of the most common causes of manufacturing quality problems — and one of the most frequently missed. Unlike a sudden failure, which triggers an immediate alarm, drift is gradual. Results stay within specification. Equipment appears to be functioning normally. And then, one day, a batch fails.

The investigation that follows almost always reveals the same thing: the drift was visible in the data for weeks or months before the failure. It just was not being looked for.

What process drift looks like

Drift is a sustained, directional change in a process output over time. It is distinct from random variation (which has no direction) and from a step change (which is sudden rather than gradual).

The most common causes of drift are:

Equipment wear. Pump seals, mixing blades, heat exchanger surfaces, and filter media all degrade over time. The degradation is gradual, and its effect on process outputs is gradual. A pump that is slowly losing efficiency will produce a gradual reduction in flow rate, which may manifest as a gradual change in product properties.

Raw material changes. Suppliers change their processes, their sources, or their specifications — sometimes without notifying customers. A raw material that is slowly changing in composition will produce a gradual change in product properties, even when the incoming material continues to pass specification.

Environmental changes. Temperature, humidity, and ambient conditions affect many manufacturing processes. Seasonal changes in ambient temperature can produce gradual changes in process outputs that are easy to mistake for random variation.

Calibration drift. Instruments that are not regularly calibrated drift over time. A temperature sensor that reads 2°C high, a flow meter that reads 3% low — these produce systematic errors in process control that manifest as gradual changes in product properties.

The signals to watch for

Drift produces characteristic patterns in process data. Learning to recognise these patterns is the foundation of effective process monitoring.

A trend in the process output. The most obvious signal: a sequence of results that is consistently moving in one direction. Six or more consecutive results that are each higher (or lower) than the previous one is a statistically significant trend, even if every individual result is within specification.

A run on one side of the average. Eight or more consecutive results that are all above (or all below) the process average is a signal that the process average has shifted. This is the most sensitive detection rule for a step change that has already occurred, and for the early stages of a drift.

Increasing distance from the centreline. Results that are consistently near the edge of the specification range, rather than near the centre, suggest that the process average has moved. This is particularly important when the results are still within specification — the margin has been reduced, and the next batch may not be.

Increasing variability. Results that are becoming more spread out over time suggest that the process is becoming less consistent. This is often a sign of equipment wear or a raw material that is becoming more variable.

Why routine monitoring misses drift

Most manufacturing quality systems are designed to detect failures, not drift. The monitoring question is: "Did this batch pass?" not "Is the process behaving consistently?"

Pass/fail monitoring answers the first question. It does not answer the second. A process can be drifting steadily toward the specification limit while every batch continues to pass — until the drift carries a batch over the limit.

Statistical process control (SPC) is designed to answer the second question. Control charts plot process outputs over time and apply detection rules that are sensitive to patterns, not just extreme values. A control chart would have flagged the drift weeks before the batch failure.

The challenge for many manufacturers is that SPC requires a systematic approach to data collection and analysis that goes beyond routine pass/fail checking. Results need to be plotted in time order, control limits need to be calculated from historical data, and someone needs to review the charts regularly and investigate signals.

A practical approach to drift detection

If you are not currently using control charts, the simplest starting point is a run chart: a time-series plot of your key process outputs, plotted in the order they were produced.

A run chart requires no statistical calculation. You plot each result as it comes in, and you look for trends and runs. It is not as sensitive as a control chart, but it is far more sensitive than pass/fail checking, and it can be set up in a spreadsheet in an hour.

Once you have a run chart, the next step is to add control limits. Calculate the mean and standard deviation of your historical results (from a period of known good performance), and add lines at the mean ± 3 standard deviations. These are your control limits. Any result outside these limits is a signal that requires investigation.

What to do when you detect drift

When a drift signal is detected, the response should be systematic:

1. Confirm the signal. Is the trend real, or is it an artefact of the measurement system? Check the calibration of the instruments used to measure the output.

2. Identify the timeframe. When did the drift start? Look at the data and identify the point at which the results began to move directionally.

3. Identify what changed. What changed in the process around the time the drift started? Raw material lots, equipment maintenance, personnel changes, environmental conditions?

4. Investigate the most likely causes. Based on the timeframe and the direction of the drift, identify the most likely causes and investigate them systematically.

5. Implement a correction. Once the cause is identified, implement a correction and monitor the process to confirm that the drift has stopped.

6. Document the investigation. Record what was found, what was done, and what the outcome was. This documentation is valuable for future investigations and for demonstrating process control to customers and regulators.

The value of early detection

The cost of a quality failure — rework, scrap, customer complaints, regulatory action — is almost always much higher than the cost of the investigation and correction that would have prevented it.

Early drift detection, through systematic monitoring and analysis, converts expensive failures into routine maintenance events. The process is corrected before the batch fails, the customer is never affected, and the investigation is straightforward because the cause is recent and identifiable.

This is the practical value of statistical process control: not the charts themselves, but the early warning they provide and the systematic response they enable.